Category: 3. Business

Introduction

Globally, hepatocellular carcinoma (HCC) remains a leading cause of cancer-related mortality, with increasing incidence projected over the next two decades. In China, it is the second most lethal cancer, and the long-term prognosis following curative surgery remains poor due to high recurrence rates, with 5-year recurrence rates reported at approximately 50%–70%.^1–6 Recurrence is most frequent within the first postoperative year and often appears as distant intrahepatic or extrahepatic metastases, thought to originate from undetected micrometastases present at surgery. Tumor-related features such as size, number, differentiation, microvascular invasion (MVI), and elevated alpha-fetoprotein (AFP) are major determinants of early recurrence. Conversely, factors including patient age, sex, underlying liver disease etiology, and cirrhosis are more commonly linked to late recurrence.^7–9 While transcatheter arterial chemoembolization (TACE) is commonly used in this setting, the introduction of immune checkpoint inhibitors (ICIs) offers a novel therapeutic option. Building on recent trials investigating ICIs in advanced or unresectable HCC, such as IMbrave150 and CheckMate 459, attention has shifted to their potential in earlier disease stages. However, limited data exist regarding their application following surgery, especially in patients with MVI. Moreover, the duration of ICI therapy necessary to maximize efficacy remains undefined. This study sought to compare ICI-based therapy with TACE in a real-world postoperative setting and to assess whether extended ICI treatment (≥12 months) offers added benefit in reducing recurrence.

Patients and Methods

Patients and Study Design

To evaluate the comparative efficacy of adjuvant ICI therapy versus TACE in improving RFS among HCC patients with MVI, we conducted a retrospective, multicenter cohort study enrolling patients who underwent curative-intent hepatic resection between January 1, 2017, and March 31, 2024, at Peking Union Medical College Hospital (Beijing), China-Japan Friendship Hospital (Beijing), and Sun Yat-sen University Cancer Center (Guangzhou). Inclusion criteria were: (1) complete (R0) resection; (2) age between 18 and 75 years; (3) pathologically confirmed HCC with MVI; and (4) Eastern Cooperative Oncology Group performance status (ECOG PS) score ≤ 1. Exclusion criteria encompassed: (1) recurrent HCC following prior curative resection; (2) history of spontaneous tumor rupture with hemorrhage; (3) coexistence of other active malignancies, including those in sustained remission; (4) administration of any neoadjuvant treatment modalities for HCC, including TACE, molecular targeted therapy, immunotherapy, or radiotherapy; and (5) death from non-HCC-related causes prior to follow-up.

Treatment and Data Collection

For patients anticipated to undergo extensive hepatectomy, preoperative evaluation of hepatic functional reserve was performed using the indocyanine green (ICG) retention test at 15 minutes. Surgical procedures were standardized across centers in accordance with established protocols previously described in the literature.¹⁰

Postoperative follow-up was routinely conducted 4 to 8 weeks after hepatic resection. For patients with MVI, adjuvant therapy was recommended contingent upon satisfactory general health and absence of contraindications. Given the lack of universally accepted adjuvant standards for MVI-positive HCC, treatment decisions were made based on multidisciplinary clinical evaluation, institutional practice patterns, and physician consensus. Regimen selection followed consistent clinical principles across centers and reflected contemporary real-world management of high-risk HCC. All adjuvant therapies, when applied, were initiated within 4 to 8 weeks following surgery.

Adjuvant TACE was initiated within 4–6 weeks post-resection, involving femoral artery catheterization and infusion of chemotherapeutic agents, as per institutional protocols.

Frequently employed targeted agents included lenvatinib,¹¹ apatinib,¹² donafenib, regorafenib,¹³ bevacizumab,¹⁴ administered once daily with dosage adjusted according to body weight or manufacturer guidelines. Immunotherapeutic agents commonly utilized in the adjuvant setting included sintilimab,¹⁵ carrelizumab,¹⁶ atezolizumab,¹⁴ tislelizumab.¹⁷ Prior to initiation, patients underwent thorough pre-treatment screening comprising blood panels, thyroid function assessment, electrocardiography, and chest CT imaging to ensure eligibility. ICIs were administered intravenously every three weeks in accordance with recommended dosing protocols. Vital signs were closely monitored during infusion and for at least 1hour post-infusion to promptly detect and manage any infusion-related reactions. Although the intended duration of adjuvant ICI therapy was 24 months, actual treatment durations varied in clinical practice. Some patients discontinued treatment within the first few months due to adverse events, financial burden, or poor adherence, even in the absence of tumor recurrence. To reduce immortal time bias in survival comparisons, patients who received ICI therapy for less than 6 months or experienced disease recurrence within 6 months were excluded from the duration-based survival analysis.

Follow-Up

Postoperative surveillance was rigorously implemented. All patients were scheduled for monthly follow-up visits during the first three months after hepatic resection, followed by assessments every three months for the subsequent two years, and semiannually thereafter. Each follow-up visits included evaluation of serum tumor markers, abdominal ultrasonography, and contrast-enhanced abdominal CT or MRI. Additional investigations, such as chest CT, bone scan, or positron emission tomography–CT (PET-CT), were performed when distant metastasis was clinically suspected. Follow-up was continued until patient death or loss to follow-up. The endpoint of the follow-up was September 20, 2024. Treatment of recurrence was personalized according to tumor profile, organ function, and patient status. The primary endpoint was RFS, as it reflects the direct effect of adjuvant therapy on preventing early relapse, particularly relevant in MVI-positive patients. OS was designated as a secondary endpoint, acknowledging the variability introduced by post-recurrence treatment heterogeneity. Of the 319 patients who received postoperative adjuvant therapy, 30 (9.4%) were lost to follow-up. Among the 80 patients under routine surveillance who did not receive adjuvant treatment, 9 (11.2%) were lost to follow-up. These individuals were not excluded from the analysis; instead, the time of their last documented follow-up was incorporated as censored observations in the survival analysis. This censoring strategy, implemented via the Kaplan–Meier method, ensured the inclusion of all available patient data and upheld the statistical robustness and validity of survival estimations.

Statistical Analysis

Continuous variables were summarized as mean ± standard deviation or median with interquartile range (IQR), and compared using Student’s t test or Mann–Whitney U-test, as appropriate. Categorical variables were compared using Pearson’s chi-square or Fisher’s exact test. To adjust for baseline differences between the TACE and immunotherapy groups, 1:1 PSM was performed using a nearest-neighbor algorithm (caliper = 0.2, no replacement). Covariates included AFP grade, HBV DNA, tumor differentiation, surgical margin, tumor number and size, satellite nodules, tumor embolus, MVI, and liver cirrhosis, selected based on clinical relevance and potential influence on treatment assignment. RFS and OS were estimated by Kaplan–Meier analysis and compared using Log rank tests; Cox regression identified independent predictors. Variables with P < 0.05 in univariate analysis, as well as clinically relevant covariates, were entered into the multivariate Cox regression model. To reduce immortal time bias, a 6-month landmark analysis excluded patients with recurrence or death before this point. Analyses were conducted using R (v4.3.1), with p < 0.05 considered significant.

Results

Patient Characteristics

From January 2017 to March 2024, a total of 1526 HCC patients from the three aforementioned centers were initially enrolled. All patients underwent curative liver resection, and postoperative pathology confirmed the diagnosis of HCC. Among them, 1,048 patients without MVI, 46 patients who received neoadjuvant therapy, 23 patients diagnosed with concurrent malignancies, and 10 patients who died from non-HCC-related causes were excluded. Ultimately, 399 patients were included in the final analysis (Figure 1). Among these patients, 132 received TACE alone, 58 received TACE combined with targeted therapy, 68 received TACE combined with targeted immunotherapy, 21 received targeted therapy combined with immunotherapy, 40 received only immunotherapy, and the remaining 80 patients did not receive any form of postoperative adjuvant therapy. Among the 319 patients who received postoperative adjuvant therapy, 30 were lost to follow-up, and among the 80 patients under active monitoring, 9 were lost to follow-up.

Figure 1 Patients flow chart. Abbreviations: MVI, microvascular invasion; HCC, hepatocellular carcinoma; TACE, transcatheter arterial chemoembolization; ICI, immune checkpoint inhibitor.

A total of 129 patients received adjuvant immunotherapy following curative hepatic resection. Among them, 40 received tislelizumab (anti–PD-1), 24 received camrelizumab (anti–PD-1), 21 received sintilimab (anti–PD-1), 11 received atezolizumab (anti–PD-L1), 10 received cadonilimab (a bispecific PD-1/CTLA-4 antibody), 10 received toripalimab (anti–PD-1), 7 received envafolimab (anti–PD-L1), and 6 received pembrolizumab (anti–PD-1). To minimize immortal time bias in the treatment duration analysis, 45 patients were excluded due to tumor recurrence occurring between 2 and 6 months postoperatively or receipt of immunotherapy for fewer than 6 months. The remaining 84 patients were included in the final analysis: 46 received adjuvant immunotherapy for less than 12 months, and 38 received it for 12 months or longer (Figure 1).

Before matching, the TACE and immunotherapy groups were largely comparable in baseline demographics and disease characteristics, except for differences in age, HBV DNA, and AFP levels. After 1:1 propensity score matching, 108 patients were included in each group (Figure 1), with no significant differences observed in baseline characteristics. A summary of baseline characteristics before and after matching is presented in Table 1.

Table 1 Baseline Characteristics of HCC Patients in the TACE and Immunotherapy Groups Before and After PSM

Treatment and Efficacy

During a median follow-up of 18 months (IQR 10–29 months), recurrence or metastasis occurred in 44 patients (34.1%) in the immunotherapy cohort and 86 patients (65.2%) in the TACE cohort. The predominant recurrence sites were the liver, lungs, and bones.

The median RFS was significantly longer in the immunotherapy cohort at 35 months (95% CI, 19–NA), compared with 16 months (95% CI, 10.8–27) in the TACE cohort (HR = 0.50, 95% CI, 0.34–0.72; p = 0.00015; Figure 2a). RFS rates at 12, 24, and 36 months were 75.4%, 53.1%, and 49.6% in the immunotherapy group, versus 54.5%, 42.8%, and 25.9% in the TACE group. Mortality was lower in the immunotherapy cohort (5.4%, 7 patients) compared to the TACE cohort (17.4%, 23 patients). Median overall survival (OS) was not reached in either group, but OS was significantly improved with immunotherapy (HR = 0.34, 95% CI, 0.14–0.80; p = 0.0096; Figure 2b). OS rates at 12, 24, and 36 months were 100.0%, 93.6%, and 86.9% in the immunotherapy cohort, versus 88.6%, 82.8%, and 81.0% in the TACE cohort. After PSM, the RFS benefit associated with immunotherapy remained statistically significant (HR = 0.54, 95% CI, 0.36–0.82; p = 0.0042; Figure 2c), while the OS difference was no longer statistically significant (HR = 0.43, 95% CI, 0.18–1.40; p = 0.060; Figure 2d).

Figure 2 Kaplan-Meier survival curves comparing the adjuvant ICI cohort and the TACE cohort: (a) RFS and (b) OS before PSM; (c) RFS and (d) OS after PSM. Abbreviations: HR, hazard ratio; CI, confidence interval.

Univariate and Multivariate Analyses of RFS and OS

Univariate and multivariate Cox regression analyses identified several independent predictors of poor prognosis (Table 2). For RFS, an advanced CNLC stage was significantly associated with shorter survival. For OS, a resection margin of less than 0.5 cm was identified as independent adverse prognostic factors.

Table 2 Univariate and Multivariate Analysis for RFS and OS of HCC Patients

Adverse Events (AE) in the Immunotherapy Cohort

Among the 129 patients who received immunotherapy, 55 patients (42.6%) experienced at least one treatment-related AE. Grade 1–2 AEs occurred in 51 patients (39.5%), while 13 patients (10.1%) experienced grade 3–4 events. No grade 5 AEs were reported (Supplementary Table 1). The most frequent AEs (any grade) were rash (9.3%), elevated AST (8.5%), ALT (7.8%), thrombocytopenia (7.0%), hypoalbuminemia (6.2%), and hypothyroidism (6.2%). Most events were grade 1–2 in severity. The most common grade 3–4 AEs included rash (5.4%), hypothyroidism (4.7%), and hypertension (3.9%). Other observed toxicities such as increased bilirubin (3.9%), elevated creatinine (3.9%), leukopenia (3.1%), mouth ulcers (3.1%), and fatigue (1.6%) were generally mild and manageable. No treatment-related deaths were observed.

Efficacy of Immunotherapy Duration on Survival Outcomes and AEs

To minimize immortal time bias, we excluded patients with recurrence-free survival less than 6 months and those who received ICI therapy for fewer than 6 months. Among the remaining cohort, patients who received adjuvant ICI therapy for 12 months or longer demonstrated significantly better RFS compared to those treated for less than 12 months (HR: 0.46, 95% CI: 0.21–0.99, p = 0.041; Figure 3a). A similar trend toward improved overall survival was observed, although the difference did not reach statistical significance (HR: 0.19, 95% CI: 0.02–1.59, p = 0.086; Figure 3b).

Figure 3 Survival outcomes of patients receiving adjuvant ICI therapy for ≥12 months versus <12 months: (a) RFS and (b) OS. Abbreviations: HR, hazard ratio; CI, confidence interval.

Among the 129 patients who received ICI-based adjuvant therapy, 68 received TACE combined with targeted immunotherapy, 21 received targeted immunotherapy, and 40 received immunotherapy alone. To account for the potential influence of different treatment regimens on survival outcomes, we conducted a subgroup analysis focusing on the largest group—patients who received postoperative TACE combined with targeted immunotherapy. After excluding those with a DFS less than 6 months and those who received immunotherapy for less than 6 months, we reevaluated RFS and OS. The analysis showed that patients who received ICIs for 12 months or longer had significantly improved RFS compared to those treated for less than 12 months (HR: 0.29, 95% CI: 0.11–0.79, p = 0.011; Figure 4a). While a longer OS was also observed in patients treated for 12 months or more, this difference was not statistically significant (HR: 0.19, 95% CI: 0.02–1.61, p = 0.089; Figure 4b).

Figure 4 Survival outcomes in the subgroup receiving TACE combined with targeted immunotherapy, stratified by ICI treatment duration (≥12 months vs <12 months): (a) RFS and (b) OS. Abbreviations: HR, hazard ratio; CI, confidence interval.

Among patients who received adjuvant immunotherapy, no significant differences were observed in the total number of adverse events between those treated for ≥12 months and those treated for <12 months (84.8% vs 89.5%, p >0.999; Supplementary Table 2). Similarly, the occurrence of grade 3–4 adverse events did not differ significantly between the two groups (26.1% vs 15.9%, p = 0.149; Supplementary Table 2). When comparing specific adverse events, no individual AE type showed a statistically significant difference between the two groups. However, numerically higher rates of hypertension and thrombocytopenia were noted in the ≥12-month treatment group, suggesting a trend toward increased incidence of some events with prolonged immunotherapy. Overall, extended treatment duration was not associated with a significantly increased risk of severe toxicity.

Discussion

Adjuvant therapies such as TACE, targeted agents, and immunotherapy have been associated with improved RFS and OS in HCC patients after curative resection.^18–20 However, the optimal duration of adjuvant immunotherapy for HCC patients with high-risk recurrence factors remains undefined, and real-world evidence on this topic is lacking. Although certain guidelines recommend that adjuvant immunotherapy should not exceed one year, they do not specify a minimum or preferred treatment duration.^21,22 In our study, compared to TACE alone, immunotherapy—either as monotherapy or in combination with TACE or targeted agents—was associated with a significant reduction in recurrence and improvement in OS among HCC patients with MVI who underwent R0 resection. After PSM, the benefit in RFS remained statistically significant, while the difference in overall survival was attenuated and no longer reached statistical significance. Importantly, ICI-based adjuvant therapies did not lead to a significant increase in AEs, with most AEs being grade 1–2, indicating good safety and tolerability. Among patients receiving adjuvant ICIs, a treatment duration of 12 months or longer was associated with significantly improved RFS compared to shorter durations. While a numerically favorable trend in OS was noted in the longer-duration group, the difference was not statistically significant. Therefore, no definitive conclusion regarding OS benefit can be drawn based on the current data, and this observation should be interpreted with caution. Furthermore, the total number of AEs and the incidence of grade 3–4 AEs were not significantly increased with longer treatment durations. Nevertheless, given the limited sample size of our study, larger and more comprehensive trials are needed to validate the safety and efficacy of postoperative adjuvant ICI therapy in this setting.

The clinical efficacy of ICIs was initially established in the advanced or unresectable HCC setting, as demonstrated by trials such as CheckMate 459²³ and IMbrave150.⁷ These studies showed that ICIs enhance antitumor immunity and may eradicate disseminated tumor cells. Extending this principle to earlier disease stages, adjuvant therapy aims to eliminate residual micrometastases and reduce distant recurrence, a major cause of treatment failure in MVI-positive patients. Indeed, several recent trials have tested this hypothesis. IMbrave050 reported an early RFS benefit with atezolizumab plus bevacizumab, though its updated analysis raised concerns about durability.²⁴ In contrast, a Phase II randomized controlled trial investigating adjuvant sintilimab showed a significant improvement in RFS among HCC patients with MVI,¹⁹ aligning with our findings and underscoring that high-risk populations may derive the greatest benefit. Retrospective studies have also suggested that adjuvant ICIs may improve prognosis among patients at high risk of recurrence.^25,26

Although several studies have investigated the safety and efficacy of adjuvant immunotherapy for HCC, the optimal duration of treatment has not been thoroughly explored. Given that treatment duration may critically influence patient outcomes, there is an urgent need for dedicated clinical trials addressing this issue. However, research specifically focused on treatment duration remains scarce. A prospective, multicenter cohort study evaluated the impact of adjuvant ICI treatment duration on RFS and OS in HCC patients at high risk of recurrence.²⁷ The results suggested that patients receiving adjuvant ICI therapy for more than six months tended to achieve better RFS and OS compared to those treated for six months or less, although the differences did not reach statistical significance. Despite the absence of a positive finding, the study indicated that six months of adjuvant ICI therapy might be insufficient and that extended treatment duration could potentially yield greater clinical benefits. Importantly, the design of ongoing Phase III randomized trials also reflects this rationale. Major studies such as CheckMate-9DX, KEYNOTE-937, JUPITER-04, SHR-1210-III-325, EMERALD-2, and DaDaLi have all adopted a 12-month adjuvant ICI regimen as the standard duration,²⁷ underscoring the clinical plausibility of our chosen cutoff. Nevertheless, the optimal duration of adjuvant immunotherapy remains an unresolved issue, not only in HCC but also in other malignancies such as non-small cell lung cancer^28,29 and melanoma, where prolonged ICI therapy has shown improved outcomes in certain settings. Our findings suggest that extending ICI therapy beyond 12 months may confer additional benefits for high-risk HCC patients; however, this hypothesis requires validation in prospective randomized studies. Future research should focus on defining the optimal duration of adjuvant immunotherapy, identifying predictive biomarkers for treatment benefit, and developing combination strategies tailored to individual recurrence risk profiles.

As a retrospective study, our analysis is inevitably subject to inherent biases. We acknowledge that comparisons based on treatment duration are vulnerable to immortal time bias, as longer-lived patients may be more likely to receive prolonged therapy. To mitigate this issue, we excluded patients who experienced recurrence or death within 6 months after surgery and those who received ICI therapy for less than 6 months. By restricting the analysis to patients who survived at least 6 months and initiated ICI treatment early, the impact of immortal time bias was reduced, although residual confounding remains possible. We also recognize that analyses involving secondary endpoints and subgroup comparisons may increase the risk of type I error. Another limitation is that the majority of ICIs used in our cohort were PD-1 inhibitors, with only a minority of patients treated with a bispecific PD-1/CTLA-4 antibody. Consequently, the efficacy and safety of other classes of immunotherapeutic agents, such as PD-L1 inhibitors, CTLA-4 inhibitors, and dual-targeting antibodies, were not assessed and warrant further investigation. Moreover, the heterogeneity of treatment regimens within our cohort may have influenced the outcomes. In addition, most patients in our study had HBV-related HCC, reflecting the epidemiological profile of HBV-endemic regions. Therefore, the generalizability of our findings to populations with HCV-related, alcohol-related, or non-viral HCC remains uncertain. Taken together, these limitations indicate that our conclusions should be interpreted with caution. Nonetheless, our findings provide important insights into the potential inadequacy of immunotherapy durations shorter than one year in high-risk HCC patients and highlight the need for prospective, standardized studies across diverse patient populations to confirm these observations.

Conclusions

This retrospective cohort study suggests that adjuvant ICI therapy following curative resection may improve RFS in HCC patients at high risk of recurrence compared to TACE. Notably, our findings indicate that a treatment duration of 12 months or longer is associated with improved RFS in patients with MVI. However, no statistically significant improvement in OS was observed with longer treatment duration. These results highlight the need to reconsider adjuvant immunotherapy strategies in this population, and underscore the importance of prospective, randomized, and large-scale clinical trials to determine the optimal duration of adjuvant ICI therapy for HCC.

Abbreviations

HCC, Hepatocellular carcinoma; MVI, Microvascular invasion; ICI, Immune checkpoint inhibitor; TACE, Transcatheter arterial chemoembolization; RFS, Recurrence-free survival; OS, Overall survival; PSM, Propensity score matching; AFP, Alpha-fetoprotein; HAIC, Hepatic arterial infusion chemotherapy; ECOG PS, Eastern Cooperative Oncology Group performance status; CT, Computed tomography; MRI, Magnetic resonance imaging; ICG, Indocyanine green; AST, Aspartate aminotransferase; ALT, Alanine aminotransferase; CNLC, China Liver Cancer staging system; BCLC, Barcelona Clinic Liver Cancer staging; PET-CT, Positron emission tomography-computed tomography; HR, Hazard ratio; CI, Confidence interval; AE, Adverse event; PD-1, Programmed death-1; PD-L1, Programmed death-ligand 1; CTLA-4, Cytotoxic T-lymphocyte-associated protein 4.

Data Sharing Statement

All data supporting the results of the study can be found in the article. Further inquiries can be directed to the corresponding author.

Statement of Ethics

This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Peking Union Medical College Hospital (Approval No. I-23PJ964). Informed consent was obtained from all individual participants included in the study.

Acknowledgments

Xiaokun Chen, Jiali Xing, and Baoluhe Zhang are co-first authors for this study. We thank all the patients and the medical staff.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Funding

This work was supported by the National Natural Science Foundation of China (81972698); the CAMS Innovation Fund for Medical Sciences (CIFMS 2021-I2M-01-014); Changzhou Xi Tai Hu development foundation for frontier cell- therapeutic technology (2024-P-019); the 2024 PhD Short-term Academic Visiting Program of Peking Union Medical College; the Start-up Fund from the Department of Liver Surgery, Peking Union Medical College Hospital; the Central high-level hospital clinical research special key cultivation project (2022-PUMCH-C-047); and 2021 Liver Cancer Diagnosis and Treatment Exchange Fund of Hubei Chen Xiaoping Science and Technology Development Foundation (CXPJJH1200009-01).

Disclosure

The authors have no conflicts of interest to declare in this work.

References

1. Han B, Zheng R, Zeng H, et al. Cancer incidence and mortality in China, 2022. J Natl Cancer Cent. 2024;4(1):47–53. doi:10.1016/j.jncc.2024.01.006

2. Rumgay H, Arnold M, Ferlay J, et al. Global burden of primary liver cancer in 2020 and predictions to 2040. J Hepatol. 2022;77(6):1598–1606. doi:10.1016/j.jhep.2022.08.021

3. Zhou L, Wang S-B, Chen S-G, et al. Risk factors of recurrence and poor survival in curatively resected hepatocellular carcinoma with microvascular invasion. Adv Clin Exp Med. 2020;29(7):887–892. doi:10.17219/acem/76750

4. Lee S, Kang TW, Song KD, et al. Effect of microvascular invasion risk on early recurrence of hepatocellular carcinoma after surgery and radiofrequency ablation. Ann Surg. 2021;273(3):564–571. doi:10.1097/SLA.0000000000003268

5. Straś WA, Wasiak D, Łągiewska B, et al. Recurrence of hepatocellular carcinoma after liver transplantation: risk factors and predictive models. Ann Transplant. 2022;27:e934924. doi:10.12659/AOT.934924

6. Hirokawa F, Hayashi M, Asakuma M, et al. Risk factors and patterns of early recurrence after curative hepatectomy for hepatocellular carcinoma. Surg Oncol. 2016;25(1):24–29. doi:10.1016/j.suronc.2015.12.002

7. Qin S, Chen M, Cheng A-L, et al. Atezolizumab plus bevacizumab versus active surveillance in patients with resected or ablated high-risk hepatocellular carcinoma (IMbrave050): a randomised, open-label, multicentre, Phase 3 trial. Lancet. 2023;402(10415):1835–1847. doi:10.1016/S0140-6736(23)01796-8

8. Pan H, Zhou L, Cheng Z, et al. Perioperative Tislelizumab plus intensity modulated radiotherapy in resectable hepatocellular carcinoma with macrovascular invasion: a phase II trial. Nat Commun. 2024;15(1):9350. doi:10.1038/s41467-024-53704-5

9. Peng N, Mao L-F, Su J-Y, et al. The efficacy and safety of tislelizumab with or without tyrosine kinase inhibitor as adjuvant therapy in hepatocellular carcinoma with high-risk of recurrence after curative resection. Front Immunol. 2025;16:1593153. doi:10.3389/fimmu.2025.1593153

10. Shi M, Guo R-P, Lin X-J, et al. Partial hepatectomy with wide versus narrow resection margin for solitary hepatocellular carcinoma: a prospective randomized trial. Ann Surg. 2007;245(1):36–43. doi:10.1097/01.sla.0000231758.07868.71

11. Kudo M, Finn RS, Qin S, et al. Lenvatinib versus sorafenib in first-line treatment of patients with unresectable hepatocellular carcinoma: a randomised phase 3 non-inferiority trial. Lancet. 2018;391(10126):1163–1173. doi:10.1016/S0140-6736(18)30207-1

12. Qin S, Li Q, Gu S, et al. Apatinib as second-line or later therapy in patients with advanced hepatocellular carcinoma (AHELP): a multicentre, double-blind, randomised, placebo-controlled, phase 3 trial. Lancet Gastroenterol Hepatol. 2021;6(7):559–568. doi:10.1016/S2468-1253(21)00109-6

13. Bruix J, Qin S, Merle P, et al. Regorafenib for patients with hepatocellular carcinoma who progressed on sorafenib treatment (RESORCE): a randomised, double-blind, placebo-controlled, phase 3 trial. Lancet. 2017;389(10064):56–66. doi:10.1016/S0140-6736(16)32453-9

14. Zhu AX, Abbas AR, de Galarreta MR, et al. Molecular correlates of clinical response and resistance to atezolizumab in combination with bevacizumab in advanced hepatocellular carcinoma. Nat Med. 2022;28(8):1599–1611. doi:10.1038/s41591-022-01868-2

15. Ren Z, Xu J, Bai Y, et al. Sintilimab plus a bevacizumab biosimilar (IBI305) versus sorafenib in unresectable hepatocellular carcinoma (ORIENT-32): a randomised, open-label, Phase 2–3 study. Lancet Oncol. 2021;22(7):977–990. doi:10.1016/S1470-2045(21)00252-7

16. Qin S, Chan SL, Gu S, et al. Camrelizumab plus rivoceranib versus sorafenib as first-line therapy for unresectable hepatocellular carcinoma (CARES-310): a randomised, open-label, international phase 3 study. Lancet. 2023;402(10408):1133–1146. doi:10.1016/S0140-6736(23)00961-3

17. Qin S, Kudo M, Meyer T, et al. Tislelizumab vs sorafenib as first-line treatment for unresectable hepatocellular carcinoma: a phase 3 randomized clinical trial. JAMA Oncol. 2023;9(12):1651–1659. doi:10.1001/jamaoncol.2023.4003

18. Wen Y, Lu L, Mei J, et al. Hepatic arterial infusion chemotherapy vs transcatheter arterial chemoembolization as adjuvant therapy following surgery for MVI-positive hepatocellular carcinoma: a multicenter propensity score matching analysis. J Hepatocell Carcinoma. 2024;11:665–678. doi:10.2147/JHC.S453250

19. Wang K, Xiang Y-J, Yu H-M, et al. Adjuvant sintilimab in resected high-risk hepatocellular carcinoma: a randomized, controlled, phase 2 trial. Nat Med. 2024;30(3):708–715. doi:10.1038/s41591-023-02786-7

20. Chen X, Wu X, Peng W, et al. Combined TACE with targeted and immunotherapy versus TACE alone improves DFS in HCC with MVI: a multicenter propensity score matching study. J Hepatocell Carcinoma. 2025;12:561–577. doi:10.2147/JHC.S504016

21. Benson AB, D’Angelica MI, Abrams T, et al. NCCN clinical practice guidelines in Oncology. (NCCN Guidelines^®). Hepatobiliary cancers. 2024 (Version 2.2024).

22. Singal AG, Llovet JM, Yarchoan M, et al. AASLD practice guidance on prevention, diagnosis, and treatment of hepatocellular carcinoma. Hepatology. 2023;78(6):1922–1965. doi:10.1097/HEP.0000000000000466

23. Yau T, Park J-W, Finn RS, et al. Nivolumab versus sorafenib in advanced hepatocellular carcinoma (CheckMate 459): a randomised, multicentre, open-label, phase 3 trial. Lancet Oncol. 2022;23(1):77–90. doi:10.1016/S1470-2045(21)00604-5

24. Yopp A, Kudo M, Chen M, et al. LBA39 Updated efficacy and safety data from IMbrave050: phase III study of adjuvant atezolizumab (atezo) + bevacizumab (bev) vs active surveillance in patients (pts) with resected or ablated high-risk hepatocellular carcinoma (HCC). Ann Oncol. 2024;35.

25. Chen W, Hu S, Liu Z, et al. Adjuvant anti-PD-1 antibody for hepatocellular carcinoma with high recurrence risks after hepatectomy. Hepatol Int. 2023;17(2):406–416. doi:10.1007/s12072-022-10478-6

26. Yang J, Jiang S, Chen Y, et al. Adjuvant ICIs plus targeted therapies reduce HCC recurrence after hepatectomy in patients with high risk of recurrence. Curr Oncol. 2023;30(2):1708–1719. doi:10.3390/curroncol30020132

27. Su J-Y, Liu SP, Xu XL, et al. Treatment duration of adjuvant immune checkpoint inhibitors in hepatocellular carcinoma patients at high risk of recurrence after resection: a prospective, multicentric cohort study. Liver Cancer. 2024;1–18.

28. Waterhouse DM, Garon EB, Chandler J, et al. Continuous versus 1-year fixed-duration nivolumab in previously treated advanced non-small-cell lung cancer: checkMate 153. J Clin Oncol. 2020;38(33):3863–3873. doi:10.1200/JCO.20.00131

29. Zalcman G, Balmaña J, Bober SL, et al. 972O Nivolumab (Nivo) plus ipilimumab (Ipi) 6-months treatment versus continuation in patients with advanced non-small cell lung cancer (aNSCLC): results of the randomized IFCT-1701 phase III trial. Ann Oncol;2022:33. doi:10.1016/j.annonc.2022.10.004

Introduction

Severe pneumonia is a major infectious disease in elderly patients, often characterized by rapid progression, multiple organ involvement, and poor prognosis.^1,2 One of its most severe complications is acute respiratory distress syndrome (ARDS), which arises from diffuse alveolar-capillary damage and manifests as refractory hypoxemia, respiratory distress, and decreased lung compliance.^3–5 Due to age-related immune decline, impaired pulmonary reserve, and high comorbidity burden, elderly patients are particularly vulnerable to developing ARDS after pneumonia, with mortality reaching 30–50%.^6,7 Despite advances in supportive care and precision medicine approaches, early diagnosis and prognosis assessment of ARDS still rely mainly on clinical criteria, imaging, and arterial blood gases,⁸ which lack sensitivity and specificity. This underscores the urgent need for reliable molecular biomarkers to improve early identification, risk stratification, and clinical decision-making.

MicroRNAs (miRNAs) are endogenous, non-coding small RNAs that regulate gene expression at the post-transcriptional level by binding to the 3′-UTR of target mRNAs. They are involved in processes such as inflammation, oxidative stress, apoptosis, and immune regulation.^9,10 miR-27a has been implicated in multiple inflammatory and malignant diseases, where it modulates immune signaling and cellular stress responses.^11,12 In pulmonary studies, miR-27a has been linked to anti-inflammatory and antioxidant effects, suggesting its potential role as a biomarker of lung injury and prognosis.¹³

Forkhead box O3 (FOXO3), a transcription factor downstream of the PI3K/Akt pathway, plays a central role in regulating apoptosis, oxidative stress responses, and inflammatory mediator release.^14,15 Previous reports indicate its involvement in lung tissue injury and ARDS pathogenesis.^16,17 Importantly, FOXO3 is a validated target of miR-27a: downregulation of miR-27a leads to FOXO3 activation, thereby promoting inflammatory cascades and exacerbating tissue damage.¹⁸

However, evidence regarding the specific expression patterns of miR-27a and FOXO3 in elderly patients with severe pneumonia complicated with ARDS remains scarce. In particular, their correlation with oxygenation index and short-term prognosis has not been fully elucidated. Therefore, this study retrospectively analyzed elderly patients with severe pneumonia, aiming to (1) detect serum levels of miR-27a and FOXO3, (2) assess their relationship with ARDS severity and 28-day mortality, and (3) evaluate their predictive value as biomarkers for prognosis. These findings may provide a molecular basis for improved survival assessment and risk stratification in elderly ARDS patients.

Materials and Methods

Study Subjects

This was a retrospective observational study including a total of 189 elderly inpatients (aged ≥60 years) with severe pneumonia admitted to the intensive care unit of our hospital from February 2023 to October 2024. According to whether ARDS was present, patients were divided into two groups: Group A (n=114, with ARDS) and Group B (n=75, without ARDS). Additionally, 70 healthy volunteers undergoing physical examination at the physical examination center during the same period were selected as the healthy control group. Among the 114 patients in Group A, there were 70 males and 44 females, with an average age of (73.16±6.72) years; in Group B, there were 46 males and 29 females, with an average age of (74.11±6.98) years; in the control group, there were 41 males and 29 females, with an average age of (73.24±7.19) years. There were no statistically significant differences in gender and age among the three groups (P>0.05), indicating comparability.

The sample size was estimated based on a preliminary analysis of 40 patients in our institution, where the difference in serum miR-27a expression between ARDS and non-ARDS patients was approximately 0.8 standard deviations. Using a two-sided α=0.05 and power (1–β)=0.80, the minimum required sample size per group was calculated as 64 cases. Considering potential dropouts and missing data, the final enrollment exceeded this requirement, ensuring adequate statistical power for group comparisons.

This study was approved by the Liberation Army General Hospital Medical Ethics Committee (Approval No.: 2024ZZLS12) and conducted in strict accordance with the ethical principles of the Declaration of Helsinki. All participants provided informed consent and signed relevant informed consent forms.

Inclusion and Exclusion Criteria

Inclusion criteria: (1) Age ≥60 years, regardless of gender; (2) First diagnosis of severe pneumonia or severe pneumonia with ARDS; (3) Diagnosis of severe pneumonia meets the criteria of the “Chinese Expert Consensus on Clinical Practice of Severe Pneumonia in Emergency Medicine”;¹⁹ (4) Diagnosis of ARDS conforms to the “Berlin Definition of Acute Respiratory Distress Syndrome”;²⁰ (5) Complete and reliable clinical data available for analysis.

Exclusion criteria: (1) Combined with malignant tumors, tuberculosis, HIV infection, or other immunodeficiency diseases; (2) Severe hepatic or renal failure, or heart failure; (3) Complicated with pulmonary tuberculosis, COPD, congenital lung dysplasia, or other pulmonary diseases; (4) History of immunosuppressant or hormone therapy within the past 6 months; (5) Complicated with severe infection at other sites or multiple organ failure; (6) Incomplete test data or improper specimen storage; (7) Pregnant or lactating women; (8) Considered unsuitable for inclusion by the researchers, such as those with psychiatric disorders or cognitive impairment who cannot cooperate with the study procedures.

Clinical Data

The following clinical data were collected for all enrolled patients: (1) Demographic data: gender; age; body mass index (BMI); smoking history; drinking history; living alone status; (2) Pneumonia-related information: type of pneumonia (community-acquired pneumonia/hospital-acquired pneumonia); whether complicated with underlying pulmonary diseases (eg, COPD, bronchiectasis, interstitial lung disease, etc).; (3) Underlying diseases: hypertension; diabetes; coronary heart disease; chronic liver disease; (4) Laboratory test indicators (within 24 h of admission): white blood cell count (WBC); C-reactive protein (CRP); procalcitonin (PCT); serum creatinine (Scr); blood urea nitrogen (BUN); (5) Mechanical ventilation: duration of mechanical ventilation.

All clinical data were extracted from the standardized hospital electronic medical record system, and only baseline information within 24 hours of admission was included to minimize bias. Data collection was independently performed by two trained researchers in a blinded manner, and any inconsistent entries were rechecked by a third investigator to ensure accuracy and reliability.

Grouping Method

Patients in Group A (with ARDS) were further subdivided according to the oxygenation index [arterial oxygen partial pressure (PaO₂) / fraction of inspired oxygen (FiO₂)] after admission: Mild ARDS subgroup: PaO₂/FiO₂ between 200–300 mmHg (n=28); Moderate ARDS subgroup: PaO₂/FiO₂ between 100–200 mmHg (n=36); Severe ARDS subgroup: PaO₂/FiO₂ <100 mmHg (n=50). Meanwhile, Group A patients were also stratified by their 28-day outcome into: Survival subgroup: patients who survived within 28 days (n=79); Death subgroup: patients who died within 28 days (n=35).

Detection of Serum miR-27a and FOXO3 mRNA

All enrolled patients underwent collection of a fasting early-morning venous blood sample (5 mL) within 24 hours of admission. Blood was placed in anticoagulant-free centrifuge tubes, left to clot at room temperature for 30 minutes, and then centrifuged at 3000 rpm for 10 minutes to isolate serum. The serum was immediately aliquoted into RNase-free centrifuge tubes (free of RNA contamination) and stored at –80°C in an ultra-low temperature freezer to ensure the quality and stability for subsequent RNA extraction. To detect the expression levels of miR-27a and FOXO3 mRNA in serum, qRT-PCR (quantitative real-time polymerase chain reaction) was used. The detection process included three main steps: total RNA extraction, reverse transcription, and real-time PCR amplification, detailed as follows: (1) Total RNA extraction: According to the instructions of the RNA extraction kit produced by Nanjing Vazyme Biotech Co., Ltd. (Product No.: RC112-01), total RNA was extracted from frozen serum. Strict RNase-free procedures were followed throughout. The concentration and purity of RNA samples were assessed using a NanoDrop™ UV spectrophotometer, and samples with A260/A280 between 1.8 and 2.1 were considered qualified. (2) Reverse transcription: RNA samples that passed quality assessment were reverse-transcribed into cDNA using the reverse transcription kit provided by Nanjing Saihongrui Biotech Co., Ltd. (Product No.: DV807A). Specific stem-loop primers were used for miRNA reverse transcription with U6 small nuclear RNA as the internal control, while mRNA reverse transcription was performed using a mixed system of Oligo(dT) and random primers, with GAPDH as the reference gene.(3) Real-time quantitative PCR amplification: The PCR reaction system was constructed based on the kit provided by Yeasen Biotechnology (Shanghai) Co., Ltd. (Product No.: 11203ES03). The total volume per reaction was 20 μL, including: 2× Master Mix buffer 10 μL, forward primer 0.5 μL, reverse primer 0.5 μL, cDNA template 2.0 μL, and RNase-free DEPC-treated water to 20 μL. Amplification was performed on an ABI 7500 real-time PCR system. Reaction conditions were: Pre-denaturation: 95°C for 30 seconds, 1 cycle; Amplification: 95°C for 5 seconds and 60°C for 40 seconds, 40 cycles; Melting curve: fluorescence signals were acquired from 65°C to 95°C, with 0.5°C increments to verify amplification specificity. (4) Data analysis and quality control: Fluorescence signal data were collected and calculated using the instrument’s built-in software. All raw amplification plots and melting curves were manually reviewed to confirm specificity. Samples with ambiguous amplification results were repeated to ensure reproducibility. Relative expression levels of miR-27a and FOXO3 mRNA were calculated using the 2^–ΔΔCt method, with U6 and GAPDH as internal controls for normalization, respectively. All primers were designed based on published sequences, verified in the NCBI database, and synthesized with quality certification by Wuhan GeneCreate Bioengineering Co., Ltd., and the primer sequences are shown in Table 1.

Table 1 Primer Sequence Information

Statistical Analysis

Statistical analysis was performed using SPSS 26.0 software, and figures were generated using GraphPad Prism 9.0. Continuous variables conforming to normal distribution were expressed as (); comparisons between two groups were made using the t-test, and comparisons among multiple groups were performed using analysis of variance (ANOVA). Categorical data were expressed as counts (n) and percentages (%), and comparisons were made using the χ²-test. Correlation analysis was performed using Pearson or Spearman correlation methods. Multivariate analysis used a binary logistic regression model to identify independent risk factors for 28-day mortality. ROC curves were plotted to compare the predictive performance of miR-27a, FOXO3, and their combination for patient mortality. Differences in AUC were compared using the Z-test. The significance level was set at α=0.05, with P<0.05 considered statistically significant.

Results

Comparison of Serum miR-27a and FOXO3 mRNA Levels Among the Three Groups

Serum miR-27a levels were significantly higher in Group B than in Group A, and further elevated in the control group compared with Group B. Conversely, FOXO3 mRNA levels were significantly lower in Group B than in Group A, and further reduced in the control group (F=77.352, 62.956, P<0.001), as shown in Figure 1. This indicates that compared with the control group, serum miR-27a levels were downregulated while FOXO3 mRNA levels were upregulated in the disease groups.

Figure 1 Comparison of serum miR-27a and FOXO3 mRNA levels among the three groups. Notes: *P<0.05 vs Group A; #P<0.05 vs Group B. All comparisons are made with reference to the control group.

Comparison of Serum miR-27a and FOXO3 mRNA Levels in Patients with Different Severities of Severe Pneumonia with ARDS

In patients with severe pneumonia and ARDS, serum miR-27a levels decreased progressively from the mild to moderate and severe subgroups, whereas FOXO3 mRNA levels increased in the same order (F=83.597, 111.834, P<0.001), as shown in Figure 2. These subgroups all belong to the disease group and reflect different severity classifications.

Figure 2 Comparison of serum miR-27a and FOXO3 mRNA levels in patients with different severities of severe pneumonia with ARDS. Notes: aP<0.05 vs mild subgroup; bP<0.05 vs moderate subgroup. All comparisons are relative within the disease group.

Correlation Between Serum miR-27a and FOXO3 mRNA Levels

Pearson correlation analysis showed that serum miR-27a and FOXO3 mRNA levels in elderly patients with severe pneumonia and ARDS were negatively correlated (r=–0.624, P<0.001), as shown in Figure 3.

Figure 3 Scatter plot of correlation between serum miR-27a and FOXO3 mRNA levels.

Spearman correlation analysis showed that the oxygenation index (mild=3, moderate=2, severe=1) was positively correlated with serum miR-27a levels (r=0.635, P<0.001), and negatively correlated with FOXO3 mRNA levels (r=–0.672, P<0.001), as shown in Figure 4. The oxygenation index was expressed in mmHg to ensure clarity.

Figure 4 Scatter plot of correlation between serum miR-27a, FOXO3 mRNA levels and oxygenation index (mmHg).

Comparison of Clinical Data in Patients with Different Prognoses of Severe Pneumonia with ARDS

The 28-day mortality rate in elderly patients with severe pneumonia and ARDS was 30.70% (35/114). The death subgroup had higher age, CRP, mechanical ventilation time, and FOXO3 mRNA levels, and lower oxygenation index and miR-27a levels compared to the survival subgroup (P<0.05). There were no statistically significant differences in other data (P>0.05), as shown in Table 2.

Table 2 Comparison of Clinical Data in Patients with Different Prognoses of Severe Pneumonia with ARDS

Multivariate Logistic Regression Analysis of Prognostic Factors in Elderly Patients with Severe Pneumonia and ARDS

Taking prognosis (survival=0, death=1) as the dependent variable, possible influencing factors from Table 1 were assigned as independent variables (see Table 3). A multivariate logistic regression model was established. Results showed that increased age, prolonged mechanical ventilation time, and elevated FOXO3 mRNA were independent risk factors, while increased oxygenation index and miR-27a levels were independent protective factors, as shown in Table 4.

Table 3 Variable Assignment Table

Table 4 Multivariate Logistic Regression Analysis of Prognostic Factors in Elderly Patients with Severe Pneumonia and ARDS

Predictive Value of miR-27a, FOXO3 mRNA, and Their Combination for Mortality in Elderly Patients with Severe Pneumonia and ARDS

The AUCs for serum miR-27a, FOXO3 mRNA, and their combination in predicting mortality in elderly patients with severe pneumonia and ARDS were 0.775, 0.781, and 0.867, respectively. The combined AUC was superior to each single index (Z_combined–miR-27a=2.557, P<0.05; Z_combined–FOXO3 mRNA=2.974, P<0.05), as shown in Table 5 and Figure 5.

Table 5 Predictive Value of miR-27a, FOXO3 mRNA, and Their Combination for Mortality in Elderly Patients with Severe Pneumonia and ARDS

Figure 5 ROC curves for predictive value of miR-27a, FOXO3 mRNA, and their combination in elderly patients with severe pneumonia and ARDS.

Discussion

This study focused on elderly patients with severe pneumonia complicated by ARDS, systematically analyzing the relationship between serum miR-27a and FOXO3 mRNA expression levels and disease severity and prognosis. Compared with healthy controls and patients with severe pneumonia without ARDS, ARDS patients showed significantly decreased serum miR-27a levels and markedly increased FOXO3 mRNA levels. Moreover, across different oxygenation index strata, miR-27a levels progressively declined with worsening ARDS severity, whereas FOXO3 mRNA levels increased stepwise relative to less severe subgroups. These findings indicate that these alterations are evident not only when compared with non-ARDS populations but also dynamically vary with disease progression, suggesting that both markers are closely associated with the pathophysiology of ARDS.

The observed inverse correlation between miR-27a and FOXO3 mRNA highlights a potential regulatory axis, in which miR-27a may play a protective role while FOXO3 promotes tissue damage. Previous studies^21,22 have confirmed that miR-27a regulates inflammatory, apoptotic, and oxidative stress pathways and plays crucial roles in various pulmonary diseases, including asthma, pulmonary fibrosis, and infections. Mechanistically, miR-27a may inhibit the release of pro-inflammatory mediators, limit oxidative damage, and reduce apoptosis by modulating NF-κB, TGF-β, and PI3K/Akt signaling pathways.^23–25 Downregulation of miR-27a weakens these protective effects, thereby amplifying inflammatory cascades. Conversely, persistent activation of FOXO3 exacerbates oxidative stress, induces mitochondrial dysfunction, and promotes immune imbalance.^26,27

Our results are consistent with the study by Lv et al,²⁸ who reported that downregulation of miR-27a aggravated alveolar injury in a murine ARDS model, whereas miR-27a mimic intervention effectively alleviated inflammation and tissue damage. FOXO3, on the other hand, is recognized as a transcription factor that promotes oxidative stress responses and cellular senescence. Wu et al²⁹ demonstrated that inhibition of FOXO3 could reduce alveolar epithelial apoptosis and preserve lung function. Together, these studies support the hypothesis that an imbalance between miR-27a and FOXO3 signaling contributes to the pathogenesis and progression of ARDS. Recent evidence also indicates that FOXO3 can influence macrophage polarization and T-cell differentiation, leading to immune dysregulation and impaired tissue repair.^30,31 These processes collectively create a vicious cycle of lung injury and inadequate repair, which aligns with the clinical features of refractory hypoxemia in elderly ARDS patients.

Clinically, our study found that the 28-day mortality rate among elderly ARDS patients reached 30.70%, higher than that reported for general ARDS populations,³² reflecting age-related vulnerability and the influence of comorbidities. Notably, multivariate logistic regression analysis indicated that elevated FOXO3 levels were an independent risk factor, whereas miR-27a and oxygenation index served as independent protective factors. ROC curve analysis showed that combined detection of these two markers achieved an AUC of 0.867, outperforming individual markers and providing a practical approach for risk stratification. Zhao et al³³ similarly demonstrated that multi-marker combined detection significantly improves prognostic prediction in ARDS patients. Therefore, this study expands the ARDS biomarker panel in elderly patients and validates the clinical utility of miR-27a and FOXO3.

In terms of novelty, this study has three main contributions. First, it is the first to combine the detection of miR-27a and FOXO3 mRNA in elderly ARDS patients, integrating molecular mechanisms with clinical prognostic assessment. Second, the inclusion of a relatively large cohort with stratification across ARDS severity enhances the clinical representativeness and reliability of the findings. Third, by focusing on elderly patients—a subgroup with poor outcomes that is often underrepresented in biomarker studies—this work fills a critical gap in ARDS research. These findings not only enrich current understanding but also provide a foundation for future therapeutic strategies targeting the miR-27a/FOXO3 signaling pathway.

However, several limitations should be acknowledged. First, as a single-center retrospective study, selection bias cannot be excluded, and multicenter prospective cohort studies are needed for validation. Second, only serum levels were assessed, lacking mechanistic validation in bronchoalveolar lavage fluid, lung tissue, or animal models. Moreover, miR-27a may regulate multiple targets beyond FOXO3, and FOXO3 may be influenced by other miRNAs or upstream signals; therefore, causal relationships remain to be confirmed. Functional experiments and multi-omics approaches could provide deeper insights into these interactions.

In conclusion, this study demonstrates that downregulated serum miR-27a and upregulated FOXO3 mRNA are closely associated with ARDS severity and short-term prognosis in elderly patients with severe pneumonia. Combined detection of these markers enhances predictive accuracy, providing a novel molecular basis for early identification, risk assessment, and potential therapeutic intervention. Future studies should integrate mechanistic validation and dynamic longitudinal monitoring to establish causal roles and explore their feasibility as intervention targets, ultimately advancing personalized management of ARDS.

Conclusion

The results of this study indicate that, compared with healthy controls and elderly patients with severe pneumonia without ARDS, serum miR-27a levels are significantly decreased, whereas FOXO3 mRNA levels are significantly increased in elderly patients with severe pneumonia complicated by ARDS. Within ARDS subgroups stratified by oxygenation index, miR-27a levels progressively decreased from mild to moderate to severe ARDS, while FOXO3 mRNA levels increased stepwise, highlighting their close association with disease severity. Elevated miR-27a may act as a protective factor, whereas elevated FOXO3 mRNA serves as an independent risk factor for poor short-term outcomes. Combined detection of these two biomarkers provides higher predictive efficacy for 28-day mortality than either marker alone, underscoring their potential utility for early risk stratification and clinical intervention. These findings suggest that the imbalance between miR-27a and FOXO3 is not only involved in the pathogenesis of ARDS but also has practical implications as prognostic biomarkers. Future studies should further investigate the molecular mechanisms underlying miR-27a regulation of FOXO3 and related downstream signaling pathways and validate their clinical utility in elderly ARDS populations through multicenter, large-sample prospective studies.

Funding

Project of National Clinical Research Center for Geriatric Diseases, Research on Diagnosis, Treatment and Comprehensive Prevention and Control Measures of Multidrug Resistant Pathogenic Bacteria Infections in Elderly Patients (Project Number: NCRCG-PLAGH-DX-2024002) 2. Military Health Care Project: Research on the Clinical Application Value of the Combined Evaluation Method of Pepsin, amylase and Lipid Cells in Tracheal Aspirates for the Diagnosis of Airway Aspiration (Project Number: 21BJZ24).

Disclosure

The authors report no conflicts of interest in this work.

References

1. Xu Y, Wang XY. [Clinical characters and influencing factors of perioperative pneumonia in elderly patients with hip fractures aged ≥80 years at a tertiary hospital in Beijing City]. Zhonghua Yu Fang Yi Xue Za Zhi. 2025;59(6):916–924. doi:10.3760/cma.j.cn112150-20250330-00257 Wolof

2. Kang M, Li J, Wan Q, et al. [Factors influencing the choice of endotracheal intubation and mechanical ventilation in patients with acute respiratory distress syndrome caused by viral pneumonia]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2022;34(6):586–591. doi:10.3760/cma.j.cn121430-20220607-00549 Dutch

3. Zhang J, Li Y, Li H, et al. [Acute respiratory distress syndrome caused by severe respiratory infectious diseases: clinical significance and solution of maintaining artificial airway closure]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2025;37(3):221–224. doi:10.3760/cma.j.cn121430-20240506-00404 Dutch

4. Tang R, Tang W, Wang D. [Predictive value of machine learning for in-hospital mortality for trauma-induced acute respiratory distress syndrome patients: an analysis using the data from MIMIC III]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2022;34(3):260–264. doi:10.3760/cma.j.cn121430-20211117-01741 Dutch

5. Liu YN, Ma XC. [The precision medicine for the treatment of acute respiratory distress syndrome is based on identification of phenotypes]. Zhonghua Yi Xue Za Zhi. 2024;104(15):1253–1257. doi:10.3760/cma.j.cn112137-20230928-00603 Danish

6. Wu J, Xiao H, Li X, et al. [Evaluation value of sequential organ failure assessment score for predicting the prognosis of patients with acute respiratory distress syndrome due to severe pneumonia]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2021;33(9):1057–1062. doi:10.3760/cma.j.cn121430-20210115-00076 Dutch

7. Zhang HL, Shang Y. [The value of point-of-care ultrasound in the diagnosis and management of acute respiratory distress syndrome]. Zhonghua Yi Xue Za Zhi. 2024;104(15):1225–1229. doi:10.3760/cma.j.cn112137-20230906-00407 Danish

8. Chalmers S, Khawaja A, Wieruszewski PM, et al. Diagnosis and treatment of acute pulmonary inflammation in critically ill patients: the role of inflammatory biomarkers. World J Crit Care Med. 2019;8(5):59–71. doi:10.5492/wjccm.v8.i5.74

9. Tian F, Ying H, Liao S, et al. lncRNA SNHG14 promotes the proliferation, migration, and invasion of thyroid tumour cells by regulating miR-93-5p. Zygote. 2022;30(2):183–193. doi:10.1017/S0967199421000319

10. Xing J, Chen LP, Yu WJ. [Non-small cell lung cancer-derived exosomal circular RNA circEZH2 activates fibroblasts by regulating the miR-495-3p / TPD52 axis and NF-κB pathway]. Zhonghua Zhong Liu Za Zhi. 2024;46(12):1176–1186. doi:10.3760/cma.j.cn112152-20230717-00015 Polish

11. Zhang G, Liu R, Dang X, et al. [Experimental study on improvement of osteonecrosis of femoral head with exosomes derived from miR-27a-overexpressing vascular endothelial cells]. Zhongguo Xiu Fu Chong Jian Wai Ke Za Zhi. 2021;35(3):356–365. doi:10.7507/1002-1892.202011026 Danish

12. Xie E, Lin M, Sun Z, et al. Serum miR-27a is a biomarker for the prognosis of non-small cell lung cancer patients receiving chemotherapy. Transl Cancer Res. 2021;10(7):3458–3469. doi:10.21037/tcr-20-3276

13. Fan X, Wang J, Qin T, et al. Exosome miR-27a-3p secreted from adipocytes targets ICOS to promote antitumor immunity in lung adenocarcinoma. Thorac Cancer. 2020;11(6):1453–1464. doi:10.1111/1759-7714.13411

14. Ruan P, Zheng Y, Dong Z, et al. [Research progress in the regulation of autophagy and mitochondrial homeostasis by AMPK signaling channels]. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2024;36(4):425–429. doi:10.3760/cma.j.cn121430-20230302-00132 Dutch

15. Yang F, Zhang H, Zhuo M, et al. [Effects of Peiminine on biological behavior and chemotherapy resistance of ovarian cancer by regulating the FOXO3-FOXM1 signaling axis]. Xi Bao Yu Fen Zi Mian Yi Xue Za Zhi. 2024;40(11):976–982. Wolof

16. Tao Y, Li -L-L, Liu S-H, et al. [FOXO3a signaling pathway in prostate cancer: progress in studies]. Zhonghua Nan Ke Xue. 2020;26(8):745–750. Basque

17. Wang Y, Liang H-X, Zhang C-M, et al. FOXO3/TRIM22 axis abated the antitumor effect of gemcitabine in non-small cell lung cancer via autophagy induction. Transl Cancer Res. 2020;9(2):937–948. doi:10.21037/tcr.2019.12.33

18. Zhang H, Xia J, Hu Q, et al. Long non‑coding RNA XIST promotes cerebral ischemia/reperfusion injury by modulating miR‑27a‑3p/FOXO3 signaling. Mol Med Rep. 2021;24(2):566.

19. Shang WF, Chen DC. [Prone positioning ventilation therapy in acute respiratory distress syndrome: knowns and unknowns in clinical efficacy]. Zhonghua Yi Xue Za Zhi. 2024;104(15):1236–1241. doi:10.3760/cma.j.cn112137-20231012-00720 Danish

20. Yuan XY, Liu L, Qiu HB. [New 2023 global definition of acute respiratory distress syndrome: progress and limitation]. Zhonghua Yi Xue Za Zhi. 2024;104(15):1216–1220. doi:10.3760/cma.j.cn112137-20231016-00770 Danish

21. Dong M, Wang X, Li T, et al. miR-27a-3p alleviates lung transplantation-induced bronchiolitis obliterans syndrome (BOS) via suppressing Smad-mediated myofibroblast differentiation and TLR4-induced dendritic cells maturation. Transpl Immunol. 2023;78:101806. doi:10.1016/j.trim.2023.101806

22. Sun Y, Jiang R, Hu X, et al. CircGSAP alleviates pulmonary microvascular endothelial cells dysfunction in pulmonary hypertension via regulating miR-27a-3p/BMPR2 axis. Respir Res. 2022;23(1):322. doi:10.1186/s12931-022-02248-7

23. Abd-Elhakim YM, Mohamed AA-R, Khamis T, et al. Alleviative effects of green-fabricated zinc oxide nanoparticles on acrylamide-induced oxidative and inflammatory reactions in the rat stomach via modulating gastric neuroactive substances and the MiR-27a-5p/ROS/NF-κB axis. Tissue Cell. 2024;91:102574. doi:10.1016/j.tice.2024.102574

24. Ong J, Faiz A, Timens W, et al. Marked TGF-β-regulated miRNA expression changes in both COPD and control lung fibroblasts. Sci Rep. 2019;9(1):18214. doi:10.1038/s41598-019-54728-4

25. Zhao XR, Zhang Z, Gao M, et al. MicroRNA-27a-3p aggravates renal ischemia/reperfusion injury by promoting oxidative stress via targeting growth factor receptor-bound protein 2. Pharmacol Res. 2020;155:104718. doi:10.1016/j.phrs.2020.104718

26. Shen J, Wang H, Wang J-S, et al. [The relationship between MicroRNA expression profiling in imatinib-resistant cell line K562/G and potential mechanism through FOXO3/Bcl-6 signaling pathway]. Zhongguo Shi Yan Xue Ye Xue Za Zhi. 2022;30(1):107–112. doi:10.19746/j.cnki.issn.1009-2137.2022.01.017 Hausa

27. Artham S, Gao F, Verma A, et al. Endothelial stromelysin1 regulation by the forkhead box-O transcription factors is crucial in the exudative phase of acute lung injury. Pharmacol Res. 2019;141:249–263. doi:10.1016/j.phrs.2019.01.006

28. Lv X, Zhang X-Y, Zhang Q, et al. lncRNA NEAT1 aggravates sepsis-induced lung injury by regulating the miR-27a/PTEN axis. Lab Invest. 2021;101(10):1371–1381. doi:10.1038/s41374-021-00620-7

29. Wu Z, Wang Y, Lu S, et al. SIRT3 alleviates sepsis-induced acute lung injury by inhibiting pyroptosis via regulating the deacetylation of FoxO3a. Pulm Pharmacol Ther. 2023;82:102244. doi:10.1016/j.pupt.2023.102244

30. Duan XH, Li H, Lyu Y, et al. Regulation of baicalin on growth of extranodal NK/T cell lymphoma cells through FOXO3/CCL22 signaling pathway. Zhongguo Shi Yan Xue Ye Xue Za Zhi. 2023;31(3):730–738. doi:10.19746/j.cnki.issn.1009-2137.2023.03.017

31. Wang H, Wang Z. Oroxylin-A attenuates taurocholate-induced lung injury via NF-κB pathway suppression. Clin Mol Epidemiol. 2025;2:5. doi:10.53964/cme.2025005

32. Wang HF, Hu WH, Song QW, et al. [Clinical study on the relationship between the exosomes in bronchoalveolar lavage fluid and plasma and the severity of lung injury and outcome in early acute respiratory distress syndrome patients]. Zhonghua Yi Xue Za Zhi. 2022;102(13):935–941. doi:10.3760/cma.j.cn112137-20211105-02448 Danish

33. Zhao C, Li Y, Wang Q, et al. Establishment of risk prediction nomograph model for sepsis related acute respiratory distress syndrome. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2023;35(7):714–718. doi:10.3760/cma.j.cn121430-20230215-00088

Hong Kong, China, 21 October 2025 – Airbus and the Cathay Group have announced a joint investment of up to US$70 million to accelerate the development of sustainable aviation fuel (SAF) production in Asia and globally. 

The agreement was announced in Hong Kong on the sidelines of the IATA World Sustainability Symposium at a ceremony hosted by Cathay Chief Operations and Service Delivery Officer Alex McGowan and Airbus President Asia-Pacific Anand Stanley.

Under the terms of the partnership, the two companies will collaborate to identify, evaluate and invest in projects that support the scaling of SAF production towards 2030 and beyond. Projects will be assessed based on their commercial viability, technology maturity, and potential for long-term offtake. 

Scaling SAF requires deep collaboration across the value chain, including from policymakers and investors to SAF producers and customers. This co-investment agreement reflects the spirit of partnership with Airbus and Cathay teaming up to accelerate production capability for more meaningful impact. 

“SAF remains the most important lever for Cathay and the wider aviation industry to drive toward our decarbonisation goals,” said Alex McGowan, Chief Operations & Service Delivery Officer, Cathay. “This co-investment partnership with Airbus underscores our commitment to building a stronger, more scalable SAF industry. It complements our broader strategy of investing in the technologies and production capacity needed for the future, including our recent investment in the oneworld BEV SAF Fund. Meanwhile we are expanding SAF usage today through partnerships with like-minded organisations.”

“This agreement reflects the shared commitment of Airbus and Cathay to make a real difference,” said Anand Stanley, President Asia Pacific, Airbus. “The production and distribution of affordable SAF at scale requires an unprecedented cross-sectoral approach. Our partnership with Cathay is a concrete example of how we catalyse production in the most suitable locations to serve our customers.”

The joint commitment also includes collaboration to advocate for supportive SAF policies on both the supply and demand side across Asia. With the region’s strong potential in feedstock supply, production capacity, and its vibrant aviation market, Airbus and Cathay aim to leverage their global experience to help shape policies that make SAF more accessible and affordable in this part of the world.

Airbus and Cathay have a long-standing partnership dating back to 1989, when the airline signed its first order for Airbus aircraft. Today, the Cathay Group operates 86 Airbus aircraft with over 70  more on order for future delivery.    

@Cathay @Airbus #SAF

Introduction

Age-related macular degeneration (AMD) is a leading cause of irreversible vision loss among individuals aged 50 years and older worldwide.¹ The global prevalence of AMD was estimated at approximately 196 million in 2020, with projections rising to 288 million by 2040 due to increasing life expectancy and aging populations in both developed and developing regions.^1,2 In India, the burden of AMD is similarly rising: recent population-based studies have reported an overall prevalence of any AMD of 1.4–2.7%, with neovascular AMD (nAMD) accounting for a substantial proportion of vision-threatening disease.³ As life expectancies increase and lifestyles change, the number of individuals at risk for AMD in India is expected to grow significantly over the next decade, underscoring the need for effective management strategies.

Since the introduction of intravitreal anti-VEGF therapy in the mid-2000s, the management of nAMD has been revolutionized. Pegaptanib was the first approved agent, but ranibizumab (Accentrix^®, Novartis India), bevacizumab, and aflibercept (Eylea^®, Regeneron) showed superior efficacy in pivotal trials.^4–7 Ranibizumab improved or maintained vision in over 90% of eyes at one year (MARINA, ANCHOR),^4,5 while aflibercept demonstrated non-inferiority with fewer injections (VIEW 1, 2).⁶ More recently, brolucizumab (Beovu^®, Novartis)⁷ offered longer dosing intervals, though safety concerns, particularly intraocular inflammation (IOI), have limited its uptake.^7,8 Despite these advances, a significant subset of eyes exhibits persistent fluid or recurrent exudation, indicating “refractory” nAMD.^9,10 Such cases often require monthly injections with suboptimal anatomical and functional outcomes, highlighting the need for therapies that target additional pathways involved in disease pathogenesis.

One such emerging option is faricimab (Vabysmo^®, Roche/Genentech, Basel, Switzerland), a bispecific monoclonal antibody that simultaneously inhibits VEGF-A and angiopoietin-2 (Ang-2).¹¹ By dual targeting, faricimab aims not only to suppress angiogenesis but also to stabilize the retinal vasculature and reduce inflammation and vascular leakage mediated by Ang-2/Tie-2 dysregulation.¹¹ In the Phase III TENAYA and LUCERNE trials, faricimab administered every 8- or 16-weeks achieved visual and anatomical outcomes non-inferior to aflibercept dosed every 8 weeks, with a median durability of 12 weeks in treatment-naïve nAMD eyes.¹² These results suggested that dual pathway inhibition could potentially improve durability and efficacy over monotherapy. Importantly, faricimab’s distinct mechanism offers a theoretical advantage in eyes that have demonstrated suboptimal response to conventional anti-VEGF-A monotherapy.

Real-world studies of faricimab are emerging, but data specifically in refractory nAMD eyes remain limited.^13–15 A retrospective series in the United States reported that a subset of refractory nAMD eyes switched to faricimab after inadequate response to aflibercept achieved decreased central retinal thickness and stability in vision over twelve months, suggesting potential benefits in a real-world context.¹³ A study from Japan reported that while 40% of aflibercept-resistant eyes could be extended to a bimonthly regimen after switching to faricimab, 59.2% ultimately discontinued the therapy for various reasons.¹⁴ To date, no published data have described the use of faricimab in refractory nAMD patients from India, where treatment access, patient demographics, and disease characteristics may differ from Western populations. This lack of local evidence creates a knowledge gap, as socioeconomic factors, genetic predispositions, and treatment adherence patterns can influence outcomes in the Indian setting.^15,16

nAMD represents a growing public health challenge, particularly in aging populations such as India’s. While anti-VEGF agents have revolutionized the management of nAMD, a subset of eyes remains refractory to standard therapies, leading to ongoing vision loss and treatment burden. Faricimab’s dual inhibition of VEGF-A and Ang-2 offers a promising therapeutic alternative in these eyes. Given the limited real-world evidence in refractory eyes and the absence of data from India, there is a clear need to evaluate faricimab’s effectiveness in this subgroup. Our study was therefore designed to assess the anatomical and functional outcomes of faricimab in eyes with refractory nAMD in a real-world Indian cohort.

Materials and Methods

This retrospective, multicenter investigation included patients managed between January 2024 and December 2025 at two tertiary care centers in India: B B Eye Foundation, Kolkata, India and Shantilal Shanghvi Eye Institute, Mumbai, India. The protocol received ethical clearance from both institutions’ review boards (BB Eye Foundation Ethics Committee and Shantilal Shanghvi Foundation Ethics Committee). All procedures adhered to the tenets of the Declaration of Helsinki, and written informed consent was obtained from each participant for treatment and data usage.

Eligible eyes were those diagnosed with nAMD that had demonstrated a refractory response to prior anti-VEGF therapy; specifically, eyes that had received at least three consecutive monthly injections of aflibercept or brolucizumab yet continued to exhibit persistent intraretinal fluid (IRF) and/or subretinal fluid (SRF) on spectral-domain OCT. Patients were required to be 50 years or older, have a confirmed diagnosis of nAMD in the study eye, and have completed a minimum of six months of follow-up at one of the two participating centers after switching to faricimab.

Eyes were excluded if any concurrent retinal or choroidal pathology could confound the diagnosis or treatment response; for example, macular neovascularization (MNV) secondary to high myopia, inflammatory causes, and other. Additional exclusions included significant media opacities (such as dense cataract or vitreous hemorrhage) that precluded reliable OCT imaging, a history of intraocular surgery (other than uncomplicated cataract extraction) within the preceding three months, concurrent diabetic retinopathy requiring treatment, advanced glaucoma, or any other ocular condition that, in the investigator’s judgment, would interfere with outcome assessment or patient safety.

At enrollment, each patient underwent a comprehensive ophthalmic evaluation performed by fellowship-trained retina specialists. Best-corrected visual acuity (BCVA) was recorded using a Snellen chart and converted to logarithm of the minimum angle of resolution (logMAR) for analysis. Intraocular pressure was measured by Goldmann applanation tonometry. Anterior segment examination was carried out with slit-lamp biomicroscopy, and dilated fundus evaluation employed 90D and 20D lenses. SD-OCT (Cirrus HD-6000; Carl Zeiss Meditec, Dublin, CA, USA) captured macular volume scans (6×6 mm, 512×128 scans) to quantify IRF, SRF, and pigment epithelial detachment (PED).

Faricimab (6.0 mg/0.05 mL) was administered on a pro re nata (PRN) basis. Injections were performed under sterile conditions in a designated minor procedure room. After topical anesthesia (proparacaine), 5% povidone-iodine was applied to the ocular surface and periocular area. The pars plana was entered with a 30-gauge needle 3.5 mm posterior to the limbus in phakic eyes (4.0 mm in pseudophakic eyes). No routine prophylactic topical antibiotics were prescribed.

Patients were evaluated monthly for the first three months post-injection and then at physician discretion, based on disease activity. At each visit, BCVA, intraocular pressure, slit-lamp biomicroscopy, and dilated fundus examination were repeated. SD-OCT scans were acquired at every follow-up to document changes in IRF, SRF, and PED. Any unscheduled visits prompted additional assessments if patients reported new symptoms (eg, diminished vision, pain, photopsia).

The primary efficacy endpoint was the change in BCVA from baseline to the final follow-up visit. Secondary endpoints included changes in the central foveal thickness, the proportion of eyes showing complete resolution of IRF, SRF, and PED on SD-OCT. Imaging at each center was evaluated by a single experienced grader each (RB and JS). In the event of any uncertainty or discrepancy in interpretation, findings were jointly reviewed between them, and a final consensus was reached to ensure consistency in assessment.

Statistical Analysis

All statistical analyses were performed using IBM SPSS Statistics for Windows, version 23.0 (IBM Corp., Armonk, NY, USA). Continuous variables, such as BCVA and CFT, were expressed as mean ± standard deviation (SD). Changes from baseline at each follow-up visit were evaluated using paired t-tests, with significance set at P<0.05.

For categorical variables including IRF, SRF, PED, and any fluid, McNemar’s test was employed to compare paired proportions at baseline and at 6 months. A two-sided P<0.05 was considered statistically significant.

Results

A total of 24 eyes from 24 patients with refractory nAMD were included in this study. The mean age of patients was 68.1 (± 10.6) years. Prior to switching to faricimab, the eyes had received an average of 11.4 (± 9.1) anti-VEGF injections, primarily aflibercept or brolucizumab. Over the six-month study period, eyes received a mean of 2.63 ± 1.34 faricimab injections (range, 1–5) on a pro-re-nata (PRN) basis: 25.0% (n=6) of eyes received one injection, 29.2% (n=7) two injections, 16.7% (n=4) three injections, 16.7% (n=4) four injections, and 12.5% (n=3) five injections. Table 1 demonstrates the demographic characteristics and treatment profile of the study eyes.

Table 1 Demographic Characteristics and Treatment-Profile of the Study Population

Best-Corrected Visual Acuity Outcomes

The mean BCVA at baseline was 0.66 (± 0.4) logMAR. Statistically significant improvements in BCVA were noted at all follow-up time points. At 1 month, mean BCVA improved to 0.47 (± 0.34) logMAR (mean change: −0.19 [± 0.26]; P=0.0003). At 2 months, further improvement was observed (0.35 [± 0.32]; mean change: −0.3 [± 0.3]; P<0.0001), with continued gains at 3 and 6 months (0.27 [± 0.26] and 0.27 [± 0.27], respectively; mean change from baseline: −0.38 [± 0.31]; P<0.0001 for both) (Table 2).

Table 2 Changes in the Best-Corrected Visual Acuity (BCVA) and Central Foveal Thickness (CFT) in the Study Population

Anatomical Outcomes

In terms of anatomical response, the mean CFT at baseline was 471.1 (± 246.4) µm. CFT showed a statistically significant reduction at each follow-up visit: 337.3 ± 198.3 µm at 1 month (mean change −133.8 [± 133.9] µm; P<0.0001), 265.1 ± 90.7 µm at 2 months (−206.0 [± 184.2] µm; P<0.0001), 217.7 ± [41.3] µm at 3 months (−253.4 [± 205.6] µm; P<0.0001), and 209.4 [± 36] µm at 6 months (−261.7 [± 208.3] µm; P<0.0001) (Table 2).

At baseline, SRF was present in 22 of 24 eyes (91.66%), IRF in 16 eyes (66.67%), and PED in 11 eyes (45.83%). Over the six-month follow-up period, significant anatomical improvements were noted across most parameters. By month 6, SRF had completely resolved in 20 of the 22 affected eyes (90.9%), with only 2 eyes showing persistent SRF. Importantly, no new cases of SRF developed during follow-up. Similarly, IRF resolved in 14 of the 16 eyes (87.5%) in which it was initially present. The remaining two eyes exhibited persistent IRF, and no new cases were noted in the previously unaffected cohort. PED demonstrated a comparatively modest response. Of the 11 eyes with PED at baseline, 6 eyes (54.5%) showed complete resolution, while 5 eyes continued to exhibit persistent PED at the end of 6 months. There were no instances of new PED development in eyes that were initially PED-free. Using McNemar’s test for paired binary outcomes, the reduction in both SRF and IRF was found to be statistically significant (P <0.001 and P=0.0006, respectively), while the change in PED did not reach statistical significance (P=0.32). Complete resolution of fluid was noted in 20/24 eyes (83.33%) at the end of six-months, which was statistically significant (P=0.00002). Table 3 demonstrates the changes in the fluid and PED status of the study eyes.

Table 3 Proportion of Eyes with Resolution of Fluid and Pigment Epithelial Detachment (PED)

Safety Analysis

No ocular or systemic adverse events were reported during the study period.

Discussion

In this retrospective real-world analysis of 24 eyes with treatment-refractory nAMD, switching to faricimab on a PRN regimen was associated with meaningful functional and anatomical improvements over six months. Visual acuity gains were both early and sustained, with mean BCVA improving from 0.66 logMAR at baseline to 0.27 logMAR at six months. Central retinal thickness decreased steadily, accompanied by high rates of fluid resolution: over 90% of eyes with baseline SRF and nearly 88% of eyes with IRF achieved complete resolution, while PED showed more modest improvement. Overall, 83% of eyes were fluid-free at six months, and no unexpected safety issues were observed. These findings suggest that faricimab may provide anatomical stability and functional benefit in patients with chronic, previously treated nAMD under real-world conditions.

The current study adds to mounting evidence that faricimab can meaningfully improve outcomes in eyes with nAMD that have proven refractory to prior anti-VEGF therapy. In our real-world cohort of refractory nAMD eyes, conversion to faricimab was associated with significant anatomical improvements; notably reductions in retinal thickness, fluid, and PED, while visual acuity was generally maintained. These findings are consistent with previous reports, including those by Tamiya R et al,¹⁷ who observed significant anatomical improvements along with preservation of visual acuity in patients with anti-VEGF resistant nAMD. Similarly, Bantounou et al¹⁸ reported favorable anatomical outcomes and stable visual acuity, achieved with a reduced number of injections. Together, these data suggest that faricimab can resolve persistent edema that has failed to clear with other agents, even when short‐term functional gains are modest. The current study’s outcomes thus align with the emerging consensus that faricimab may rescue patients in whom prior anti-VEGF therapy has plateaued, reducing fluid burden without compromising safety.

Mechanistically, faricimab’s efficacy in this setting is readily explained by its unique dual-target action. Faricimab is a bispecific monoclonal antibody that simultaneously binds vascular endothelial growth factor A (VEGF-A) and angiopoietin-2 (Ang-2).^11,12 The VEGF pathway is the well-known driver of neovascular growth and leakage in AMD, and all prior first-line treatments (bevacizumab, ranibizumab, aflibercept, brolucizumab) target VEGF‐A or its family. Angiopoietins (primarily Ang-1 and Ang-2) regulate vascular stability via the Tie2 receptor: Ang-1/Tie2 signaling promotes quiescence and tight endothelial junctions, whereas elevated Ang-2 (usually released from hypoxic or stressed endothelium) competes with Ang-1 and effectively destabilizes vessels, making them more permeable and prone to inflammation.¹⁹ In nAMD, chronic hypoxia and inflammation drive Ang-2 upregulation, so that even if VEGF is neutralized, ongoing Ang-2–mediated permeability and inflammatory signaling can sustain fluid. By simultaneously inhibiting Ang-2 and VEGF-A, faricimab promotes vascular stabilization and mitigates inflammatory processes.^17–19 In practical application, this results in a more comprehensive inhibition of both angiogenic signaling and vascular permeability pathways. Thus, faricimab’s mechanism directly addresses a hypothesized contributor to refractory fluid: elevated Ang-2 and persistent vascular leak despite prior VEGF blockade. If a patient’s persistent edema is partly driven by Ang-2–mediated inflammation and microvascular instability, faricimab is the first available therapy that can counteract both pathogenic arms simultaneously.

Prior strategies for refractory AMD, including lateral switches among VEGF agents or the use of higher-dose or longer-acting molecules, have had variable and often incomplete success.^20,21 While newer agents like brolucizumab showed potent drying effects and extended durability in pivotal trials, concerns over intraocular inflammation (IOI) and rare but severe instances of retinal vasculitis have limited their adoption in clinical practice.^8,9 Faricimab, by targeting an additional angiogenic pathway without a marked increase in inflammatory risk, offers an appealing alternative.^17–19 We hypothesize that by targeting a complementary angiogenic pathway, one not addressed by earlier agents, faricimab underlies the superior visual and anatomical outcomes we observed, all while avoiding the immune-mediated toxicity profile characteristic of brolucizumab.

Beyond its dual-target action, faricimab offers practical advantages that are especially relevant in a high-burden setting. The pivotal TENAYA and LUCERNE trials showed that faricimab dosed up to every 16 weeks achieved non-inferior visual outcomes compared to aflibercept every 8 weeks.¹² By two years, ~60–80% of patients on faricimab could be extended to 12- or 16-week intervals.¹² This durability was mirrored in DME trials (YOSEMITE/RHINE) and small real-world studies;²² for example, Penha et al²³ report that faricimab treated patients often achieved 12-week or longer dosing schedules in practice. In our study, over six months, more than half of eyes (54.2%) required two or fewer injections after switching; 25.0% received a single injection and 29.2% received two, underscoring the potential to reduce treatment burden. In India, where adherence is often compromised by travel difficulties, cost, and comorbidities, such extended intervals can be transformative. Frequent anti-VEGF visits (4–8 week intervals) impose heavy logistic and financial strain. Indeed, even in well-resourced settings only a small minority of patients can sustain ≥12-week intervals with standard care.¹⁸ By contrast, faricimab’s protocol (with the option of Q12–16W dosing) directly addresses an unmet need in real-world management of recalcitrant nAMD, potentially improving adherence and outcomes over time.

It should be noted that global experience with faricimab in refractory AMD is still emerging. A few recent reports illustrate its promise but also highlight the need for more data, especially in diverse populations. Baek et al²⁴ found that faricimab reduced injection burden and improved visual and anatomical outcomes in eyes unresponsive to other agents. Bantounou et al¹⁸ observed that faricimab produced rapid fluid resolution and decreased injection frequency in previously treated nAMD, again with stable VA. Tamiya et al¹⁷ observed that over half of their aflibercept-refractory eyes had fluid reduction after one faricimab injection, and 25% achieved a dry macula at 2 months without recurrence for up to 4 months. These series consistently report anatomical gains with visual stabilization or improvement. However, none of these studies included substantial numbers of Indian patients. Our study is thus timely: by providing real-world data on faricimab in refractory nAMD in an Indian context, it fills a critical gap. To our knowledge, no prior published series from India has evaluated faricimab in this specific population. Given potential racial, genetic and healthcare differences, it cannot be assumed that Western findings extrapolate perfectly to Indian eyes. Our study’s population, often older patients with significant macular pathology, limited resources, and irregular follow-up, reflects “real life” conditions in India. The fact that faricimab produced clear anatomic benefits in this cohort supports its generalizability and suggests it is a viable tool in the Indian retina armamentarium.

Nonetheless, the current study has inherent limitations. As a retrospective, single-arm review, it cannot prove efficacy with the rigor of a randomized trial. There is no concurrent control group, and selection bias (which eyes were chosen for switching) likely influenced outcomes. Follow-up is relatively short, and end-points like VA are affected by ceiling/floor effects and chronic scarring in these eyes. We also did not analyze patient-reported outcomes or long-term retreatment rates. On the other hand, the study’s strengths include its multi-center design and “real-world” heterogeneity; we included patients who in practice would not meet strict trial criteria (eg very chronic lesions, multiple previous injections). The findings therefore complement the controlled trials by showing what happens in everyday clinics. Importantly, no unexpected safety issues arose: faricimab was well-tolerated, with no cases of IOI being reported.

Conclusion

In summary, the current study shows that faricimab, by neutralizing both VEGF-A and Ang-2, delivers meaningful visual and anatomical gains in refractory nAMD while addressing dual angiogenic pathways. Importantly, over half of eyes required two or fewer injections over six months, underscoring a substantial reduction in treatment burden. In an Indian context, where real-world data are limited, these findings suggest that retina specialists can expect outcomes on par with global reports. Practically, faricimab may be indicated in cases with persistent edema or suboptimal response to other anti-VEGF therapies. Although vigilance for IOI remains essential, the balance of robust efficacy and fewer injections makes faricimab a valuable switch option. However, these findings should be interpreted with caution. Larger, prospective studies with longer follow-up are needed to validate and refine retreatment strategies in this cohort.

Disclosure

J.US is affiliated with Shantilal Shanghvi Foundation (SSF), outside the submitted work. The authors declare that they have no other competing interests in this work.

References

1. Wong WL, Su X, Li X, et al. Global prevalence of age-related macular degeneration and disease burden projection for 2020 and 2040: a systematic review and meta-analysis. Lancet Glob Health. 2014;2(2):e106–e116. doi:10.1016/S2214-109X(13)70145-1

2. Sheth JU, Stewart MW, Narayanan R, et al. Macular neovascularization. Surv Ophthalmol. 2025;70(4):653–675. doi:10.1016/j.survophthal.2024.08.003

3. Hamati J, Prashanthi S, Narayanan R, et al. Prevalence of age-related macular degeneration and associated factors in Indian cohort in a tertiary care setting. Indian J Ophthalmol. 2023;71(10):3361–3366. doi:10.4103/IJO.IJO_199_23

4. Brown DM, Michels M, Kaiser PK, et al. Ranibizumab versus verteporfin photodynamic therapy for neovascular age-related macular degeneration: two-year results of the ANCHOR study. Ophthalmology. 2009;116(1):57–65.e5. doi:10.1016/j.ophtha.2008.10.018

5. Rosenfeld PJ, Brown DM, Heier JS, et al. Ranibizumab for neovascular age-related macular degeneration. N Engl J Med. 2006;355(14):1419–1431. doi:10.1056/NEJMoa054481

6. Heier JS, Brown DM, Chong V, et al. Intravitreal aflibercept (VEGF trap-eye) in wet age-related macular degeneration [published correction appears in Ophthalmology. Ophthalmology. 2012;119(12):2537–2548. doi:10.1016/j.ophtha.2012.09.006

7. Dugel PU, Koh A, Ogura Y, et al. HAWK and HARRIER: Phase 3, multicenter, randomized, double-masked trials of brolucizumab for neovascular age-related macular degeneration. Ophthalmology. 2020;127(1):72–84. doi:10.1016/j.ophtha.2019.04.017

8. Chakraborty D, Maiti A, Sheth JU, et al. Brolucizumab in neovascular age-related macular degeneration – indian real-world experience: the BRAILLE study – fifty-two-week outcomes. Clin Ophthalmol. 2022;16:4303–4313. doi:10.2147/OPTH.S395577

9. Chakraborty D, Maiti A, Sheth JU, et al. Brolucizumab in neovascular age-related macular degeneration – indian real-world experience: the BRAILLE study. Clin Ophthalmol. 2021;15:3787–3795. doi:10.2147/OPTH.S328160

10. Ashraf M, Banaee T, Silva FQ, Singh RP. Switching anti-vascular endothelial growth factors in refractory neovascular age-related macular degeneration. Ophthalmic Surg Lasers Imaging. 2018;49(3):166–170. doi:10.3928/23258160-20180221-03

11. Agostini H, Abreu F, Baumal CR, et al. Faricimab for neovascular age-related macular degeneration and diabetic macular edema: from preclinical studies to phase 3 outcomes. Graefes Arch Clin Exp Ophthalmol. 2024;262(11):3437–3451. doi:10.1007/s00417-024-06531-9

12. Khanani AM, Kotecha A, Chang A, et al. TENAYA and LUCERNE: two-year results from the phase 3 neovascular age-related macular degeneration trials of faricimab with treat-and-extend dosing in year 2. Ophthalmology. 2024;131(8):914–926. doi:10.1016/j.ophtha.2024.02.014

13. Rush RB. One-year outcomes of faricimab treatment for aflibercept-resistant neovascular age-related macular degeneration. Clin Ophthalmol. 2023;17:2201–2208. doi:10.2147/OPTH.S424315

14. Kataoka K, Itagaki K, Hashiya N, et al. Six-month outcomes of switching from aflibercept to faricimab in refractory cases of neovascular age-related macular degeneration. Graefes Arch Clin Exp Ophthalmol. 2024;262(1):43–51. doi:10.1007/s00417-023-06222-x

15. Chakraborty D, Das S, Maiti A, et al. Clinical evaluation of faricimab in real-world diabetic macular edema in India- a multicenter observational study. Clin Ophthalmol. 2025;19:269–277. doi:10.2147/OPTH.S502033

16. Soman M, Nair I, Sheth JU, Nair U. Innovator Versus Biosimilar Ranibizumab in Polypoidal Choroidal Vasculopathy: real-World Evidence. Ophthalmol Ther. 2022;11(3):1175–1186. doi:10.1007/s40123-022-00507-w

17. Tamiya R, Hata M, Tanaka A, et al. Therapeutic effects of faricimab on aflibercept-refractory age-related macular degeneration. Sci Rep. 2023;13(1):21128. doi:10.1038/s41598-023-48190-6

18. Bantounou MA, Elsheikh M, Ijasan A, Santiago C. Real-world experience of intravitreal faricimab injection in previously treated neovascular age-related macular degeneration eyes: a case series. BMC Ophthalmol. 2025;25(1):117. doi:10.1186/s12886-025-03953-9

19. Khanani AM, Russell MW, Aziz AA, et al. Angiopoietins as potential targets in management of retinal disease. Clin Ophthalmol. 2021;15:3747–3755. doi:10.2147/OPTH.S231801

20. Yiu G, Gulati S, Higgins V, et al. Factors Involved in Anti-VEGF treatment decisions for neovascular age-related macular degeneration: insights from real-world clinical practice. Clin Ophthalmol. 2024;18:1679–1690. doi:10.2147/OPTH.S461846

21. Fu Y, Zhang Z, Webster KA, Paulus YM. Treatment strategies for anti-VEGF resistance in neovascular age-related macular degeneration by targeting arteriolar choroidal neovascularization. Biomolecules. 2024;14(3):252. doi:10.3390/biom14030252

22. Wong TY, Haskova Z, Asik K, et al. Faricimab treat-and-extend for diabetic macular edema: two-year results from the randomized phase 3 YOSEMITE and RHINE trials. Ophthalmology. 2024;131(6):708–723. doi:10.1016/j.ophtha.2023.12.026

23. Penha FM, Masud M, Khanani ZA, et al. Review of real-world evidence of dual inhibition of VEGF-A and ANG-2 with faricimab in NAMD and DME. Int J Retina Vitreous. 2024;10(1):5. doi:10.1186/s40942-024-00525-9

24. Baek SC, Jeong A, Min Sagong M. Real-world efficacy of faricimab in patients with treatment-resistant neovascular age-related macular degeneration: outcomes at six months. J Retina. 2024;9:150–155. doi:10.21561/jor.2024.9.2.150

Introduction

In-hospital cardiac arrest (IHCA) remains a frequent and critical event that places a substantial emotional and operational burden on healthcare teams. Once IHCA occurs, the prognosis is poor: more than half of patients do not survive despite resuscitation, and nearly 90% of survivors suffer significant neurological impairment.¹ The sudden onset of IHCA, often following rapid but under-recognized clinical deterioration, makes early detection particularly challenging. This is especially true in general wards, where approximately 72% of IHCAs occur.^2–4 Reported survival rates vary by region, with recent US data indicating a survival-to-discharge rate of about 25.8%,^5,6 whereas a Taiwanese study showed a return of spontaneous circulation (ROSC) in 66% of cases but survival-to-discharge of only 11.8%.⁴

Although IHCA management strategies are often adapted from out-of-hospital cardiac arrest (OHCA) research, important differences exist in epidemiology and underlying pathophysiology.⁷ Conventional risk assessment methods typically rely on medical history, trends in vital signs, laboratory values, and procedural data to estimate clinical deterioration or mortality risk.⁸ However, relatively few studies have specifically focused on identifying predictors of unexpected IHCA before the event, rather than outcomes after resuscitation.

To improve early recognition, clinical scoring systems such as the National Early Warning Score (NEWS) and the Modified Early Warning Score (MEWS) are widely used, particularly in the United Kingdom.⁹ Other early warning systems, such as the Cardiac Arrest Risk Triage (CART) score,¹⁰ have also been implemented in general wards in the United States. These scores depend mainly on vital signs to identify patients at risk of acute deterioration, including cardiac arrest. Their predictive performance, however, is modest, with reported areas under the receiver operating characteristic curve (AUC) ranging from 0.65 to 0.79.¹¹

Once the high-risk patient group is identified, high-intensity care should be initiated, such as frequent vital sign monitoring, activation of rapid response teams or ICU admission for the most severe cases. According to a systematic review by Hogan et al, the implementation of the National Early Warning Score (NEWS) in daily practice-accompanied by the use of different algorithms-was associated with a 6.4% annual reduction in in-hospital cardiac arrest (IHCA) incidence and a 5% annual improvement in survival rates.¹²

The widespread adoption of electronic health records and digital healthcare systems has created opportunities for advanced predictive analytics. By leveraging dynamic, longitudinal patient data, predictive models may detect clinical deterioration earlier and with greater accuracy. Prior studies have shown that machine learning (ML) methods-such as random forest, XGBoost, decision trees, and multivariate adaptive regression splines (MARS)-often outperform traditional statistical models in predicting mortality and major cardiovascular events.^13,14 Ensemble ML approaches, which combine multiple algorithms, have demonstrated even stronger accuracy and calibration in clinical applications.¹⁵

Despite these advances, most existing studies have focused on post-arrest outcomes or on predicting OHCA, leaving a critical gap in pre-arrest risk stratification for IHCA.^16,17 Only a limited number of studies have begun to explore IHCA prediction, primarily by evaluating traditional risk factors with conventional statistical methods.^18,19

To address this, the present study compares the predictive performance of conventional logistic regression with four ML algorithms-random forest, XGBoost, decision tree, and MARS-for forecasting IHCA among hospitalized patients. By incorporating comprehensive clinical variables, this study aims to enhance early risk stratification and support proactive interventions to reduce IHCA incidence and improve patient outcomes.

Materials and Methods

We conducted a retrospective, single-center, case-control study at National Taiwan University Hospital (NTUH), including adult patients (≥18 years) who experienced unexpected in-hospital cardiac arrest (IHCA) between 2011 and 2018. Eligible patients were required to have at least one documented electrocardiogram (ECG) prior to the IHCA event. The study protocol was approved by the Institutional Review Board of NTUH (IRB No. 201807063RINC). This study was conducted in accordance with the principles of the Declaration of Helsinki. Given the retrospective design and the use of de-identified data, the need for informed consent was waived.

For the control cohort, 4,000 patients were randomly selected from 205,999 hospitalized individuals without CPR events during the study period. Patients with do-not-resuscitate (DNR) orders at admission (n = 65) or with incomplete clinical records (n = 471) were excluded, resulting in 3,464 patients in the non-IHCA group. The selection and exclusion process is shown in Figure 1. Incomplete clinical records were defined as the absence of essential demographic information (eg, age, sex, comorbidities) or more than 30% missing vital sign or laboratory variables. For the remaining dataset, variables with ≤30% missing data were imputed using multiple imputation by chained equations (MICE). The percentage of missing data for each variable is summarized in Table S1.

Figure 1 Flow diagram of study population selection. Adult inpatients at NTUH (2011–2018) with documented ECG (n = 207,290) were classified according to in-hospital CPR status. After exclusions, the IHCA group (with in-hospital CPR) comprised 800 patients and the non-IHCA group (without in-hospital CPR) comprised 3,464 patients. Abbreviations: CPR, cardiopulmonary resuscitation; DNR, do-not-resuscitate; ECG, electrocardiogram; IHCA, in-hospital cardiac arrest; NTUH, National Taiwan University Hospital.

The primary outcome was IHCA, defined as the absence of a palpable pulse with attempted resuscitation during hospitalization. The dataset included four major domains of variables. Demographic information comprised age, sex, and body mass index (BMI). Comorbidities were identified from medical records and coded using the International Classification of Diseases, Ninth and Tenth Revisions (ICD-9/10-CM). Vital signs included systolic blood pressure (SBP), diastolic blood pressure (DBP), mean blood pressure (MBP), pulse rate, respiratory rate, and body temperature. Laboratory parameters included serum creatinine, serum sodium, serum potassium, hemoglobin, platelet count, aspartate aminotransferase (AST), and alanine aminotransferase (ALT). Diagnoses were coded using the International Classification of Diseases, Ninth and Tenth Revisions (ICD-9-CM/ICD-10-CM), and procedural codes were obtained from Taiwan’s National Health Insurance execution code system.

Five predictive models were developed: logistic regression, decision tree, random forest, extreme gradient boosting (XGBoost), and multivariate adaptive regression splines (MARS). Data preprocessing included quality checks and imputation of missing values to ensure integrity. The dataset was randomly divided into training (80%) and testing (20%) subsets. Model training used 10-fold cross-validation for hyperparameter optimization and to minimize overfitting. Figure 2 illustrates the ML analytical workflow used in our study.

Figure 2 Modeling workflow. Data were processed and split into training and testing datasets. Five algorithms (LR, DT, RF, XGB, MARS) were trained on the training dataset, evaluated on standard metrics (AUC, accuracy, sensitivity, specificity, F1 score), and variable importance was summarized by average rank across models. Abbreviations: AUC, area under the curve; DT, Decision Tree; LR, Logistic Regression; MARS, Multivariate Adaptive Regression Splines; RF, Random Forest; XGB, Extreme Gradient Boosting.

Logistic regression was used as a benchmark model for binary classification, estimating the probability of IHCA based on clinical predictors. It remains widely applied in medical research and serves as a reference for comparing the performance of more advanced ML algorithms.

Decision trees are supervised learning models that classify outcomes by sequentially splitting data into subgroups based on predictor variables. Each branch represents a decision rule, and terminal nodes represent predicted outcomes. Their hierarchical, rule-based structure makes them intuitive and interpretable for both technical and clinical applications.

Random forest is an ensemble method that improves the stability and accuracy of decision trees. It generates multiple trees using bootstrap samples with randomized feature selection and aggregates their results by majority voting. Out-of-bag samples are used to estimate generalization error and feature importance, reducing overfitting and enhancing predictive reliability.

XGBoost is an optimized gradient boosting algorithm that combines multiple weak learners, typically decision trees, into a strong predictive model. It incorporates parallel processing, automated handling of missing data, and regularization to reduce overfitting. XGBoost has demonstrated state-of-the-art performance on structured clinical datasets and is widely applied in healthcare risk prediction.

Multivariate Adaptive Regression Splines (MARS) is a non-linear regression technique that models complex relationships using adaptive spline functions. It builds models through forward selection of candidate basis functions followed by backward elimination to control complexity. This flexibility allows MARS to capture both linear and non-linear effects, making it suitable for identifying subtle patterns in clinical data.

While a concise overview of each model is presented here, detailed algorithmic descriptions and hyperparameter specifications are provided in Supplementary Material 1.

To minimize the impact of potential multicollinearity among predictors (eg, renal markers, ECG intervals), we applied L1 regularization when constructing logistic regression models, which performs variable selection and shrinks the coefficients of less informative or collinear variables. For the machine learning approaches, we primarily employed tree-based models (eg, random forest, XGBoost), which are inherently less sensitive to multicollinearity due to their recursive partitioning mechanisms. Together, these strategies reduced the influence of collinearity and enhanced the robustness of our analyses.

Model performance was evaluated using standard classification metrics. Accuracy was defined as the proportion of correct predictions among all cases. Sensitivity (recall, true positive rate) represented the proportion of actual positives correctly identified, whereas specificity (true negative rate) represented the proportion of actual negatives correctly identified. Positive predictive value (PPV, precision) indicated the proportion of predicted positives that were truly positive, and negative predictive value (NPV) indicated the proportion of predicted negatives that were truly negative. The F1 score, calculated as the harmonic mean of precision and recall, provides a single measure balancing false positives and false negatives, as shown in Equation (1). Finally, the AUC summarized overall discrimination across all decision thresholds, reflecting the probability that a randomly selected positive case would be ranked higher than a randomly selected negative case (0.5 = no discrimination; 1.0 = perfect discrimination).

(1)

Equation (1). Formula for calculating the F1 score.

All analyses were performed using R software (version 4.0.3) within RStudio (version 1.4.1103), with dedicated R packages supporting each ML algorithm. Logistic regression was implemented using the glmnet package (version 4.1–1), decision trees with the rpart package (version 4.1–15), random forests with the randomForest package (version 4.6–14), and XGBoost with the xgboost package (version 1.5.0.1). MARS was conducted using the earth package (version 5.3.2). The caret package (version 6.0–90) was used for model training, hyperparameter tuning, and the evaluation of variable importance across methods.

An advanced language model (ChatGPT 5, OpenAI, San Francisco, CA, USA) was employed to enhance the grammar, phrasing, and readability of the manuscript. The model did not contribute to scientific content, data analyses, or interpretation. All generated text was thoroughly examined and edited by the authors, who assume full responsibility for the accuracy and conclusions of the manuscript.

Results

As summarized in Table 1, a total of 800 patients with IHCA and 3,464 randomly selected hospitalized controls were analyzed. Compared with controls, the IHCA group was significantly older (64.6 ± 15.9 vs 57.0 ± 16.6 years, p < 0.001), had a slightly higher proportion of males (60.4% vs 56.5%, p = 0.048), and a lower mean body mass index (23.6 ± 5.0 vs 24.3 ± 4.2 kg/m², p < 0.001).

Table 1 Comparison of Baseline Characteristics Between IHCA and Non-IHCA Groups

Cardiovascular comorbidities were markedly more prevalent in the IHCA group, including heart failure (43.2% vs 7.7%), acute coronary syndrome (ACS) (23.8% vs 3.0%), chronic coronary syndrome (42.8% vs 16.7%), peripheral artery disease (13.9% vs 4.2%), and hypertension (59.2% vs 41.1%) (all p < 0.001). Non-cardiovascular conditions such as diabetes mellitus (41.2% vs 20.5%), chronic kidney disease (32.9% vs 10.2%), and end-stage renal disease (20.4% vs 5.3%) were also more frequent (all p < 0.001). In contrast, malignancy was less common among IHCA patients (43.0% vs 50.9%, p < 0.001), although both groups demonstrated a high prevalence of malignancy.

Laboratory findings indicated greater systemic inflammation and renal dysfunction in IHCA patients, with significantly higher white blood cell counts (11.63 vs 7.29 × 10³/μL), blood urea nitrogen (BUN) (37.8 vs 17.8 mg/dL), and creatinine (2.31 vs 1.08 mg/dL) (all p < 0.001). However, liver function markers such as AST and ALT were not further analyzed because a high proportion of missing data was detected. This was likely due to local clinical practice patterns, where physicians often order only one of these tests rather than both, partly influenced by insurance-related considerations. IHCA patients also exhibited more pronounced anemia (hemoglobin 11.0 vs 13.1 g/dL) and thrombocytopenia (198.6 vs 239.9 × 10³/μL) (both p < 0.001). Serum potassium did not differ significantly. Electrocardiographic intervals were consistently prolonged, with longer ECG PR interval (151 vs 127 ms), ECG QRS duration (100 vs 90 ms), and corrected QT interval on ECG (471 vs 431 ms) (all p < 0.001).

Vital sign comparisons revealed higher pulse rates (92.9 vs 79.7 bpm, p < 0.001) and respiratory rates (20.2 vs 18.4 breaths/min, p < 0.001) among IHCA patients. Blood pressure values were slightly lower, including systolic (127.2 vs 130.2 mmHg, p < 0.001), diastolic (72.3 vs 77.2 mmHg, p < 0.001), and mean blood pressure (90.1 vs 94.4 mmHg, p < 0.001). Body temperature was minimally higher (36.46 vs 36.40°C, p = 0.006). These findings collectively indicated a profile of advanced comorbidity burden, systemic inflammation, renal dysfunction, anemia, and hemodynamic compromise in the IHCA group.

As shown in Table 2, model discrimination ranged from moderate to excellent (AUC 0.739–0.910). The decision tree performed weakest overall, with an AUC of 0.739, sensitivity of 0.331, and the lowest F1 score of 0.450, despite excellent specificity (0.965). By comparison, ensemble approaches achieved superior discrimination. Random forest yielded the highest AUC (0.910) and the strongest positive predictive value (0.749), but this improvement in precision was accompanied by reduced sensitivity (0.544). XGBoost provided the most balanced performance, with an AUC of 0.909, accuracy of 0.883, sensitivity of 0.615, specificity of 0.949, NPV of 0.914, and F1 score of 0.675, representing the highest sensitivity among all models while maintaining excellent overall accuracy. MARS also showed consistent performance across metrics (AUC 0.897; accuracy 0.881; sensitivity 0.580; specificity 0.952; F1 score 0.667), highlighting its stability and calibration.

Table 2 Performance of the LR, Decision Tree, Random Forest, XGBoost and MARS Methods

Logistic regression, although a conventional statistical approach, remained competitive. It achieved an AUC of 0.895 and accuracy of 0.876, with PPV 0.724 and NPV 0.907. However, sensitivity was only moderate (0.580). Overall, these results indicate that ensemble machine learning methods (XGBoost and random forest) outperformed single decision trees and conventional regression in terms of discriminatory power. XGBoost was the only model to achieve both high sensitivity and strong overall accuracy, while MARS provided well-balanced performance with interpretable nonlinear modeling.

Variable importance rankings are summarized in Table 3. Despite differences in methodology, there was strong convergence across models on several key predictors. Logistic regression prioritized hemoglobin, pulse rate, ACS, heart failure, and platelet count. In contrast, the machine learning models consistently ranked BUN and corrected QT interval on ECG among the top predictors, followed by hemoglobin, heart failure, and pulse rate.

Table 3 Comparative Variable Importance Rankings and Average Ranks Across Five Predictive Models

When averaged across all five models, the top predictors were BUN, corrected QT interval on ECG, hemoglobin, heart failure, pulse rate, platelet count, ACS, white blood cell count, respiratory rate, and serum sodium. These features represented multiple domains: renal dysfunction and metabolic derangement (BUN, serum creatinine, serum sodium), chronic cardiovascular comorbidities (heart failure, ACS), hematologic impairment (hemoglobin, platelet count), systemic stress and inflammation (pulse rate, respiratory rate, white blood cell count), and electrophysiological abnormalities (corrected QT interval on ECG, ECG QRS duration).

The decision tree model presented in Figure 3 further demonstrates how a limited set of key predictors can effectively stratify IHCA risk. For example, pathways incorporating thresholds for BUN (<27 mg/dL), pulse rate, and heart failure status effectively separated patients into high- and low-risk subgroups with minimal computational steps. This simplified structure underscored the consistency of these variables across different modeling approaches.

Figure 3 Decision tree model for IHCA prediction. The model stratified IHCA risk using key variables including BUN, HF, pulse rate, DBP, Hb, ACS, and ECG QTc, with terminal nodes showing predicted probabilities. Abbreviations: ACS, acute coronary syndrome; BUN, blood urea nitrogen; DBP, diastolic blood pressure; ECG QTc, corrected QT interval on ECG; Hb, hemoglobin; HF, heart failure; MBP, mean blood pressure.

Together, these results demonstrate that IHCA was associated with a multifactorial risk profile characterized by advanced age, cardiovascular comorbidities, renal dysfunction, hematologic abnormalities, and electrophysiological instability. Among the predictive models, ensemble machine learning approaches, particularly XGBoost and random forest, provided the highest discriminatory power, whereas MARS delivered stable and well-balanced performance. Logistic regression, although less powerful, remained a robust and interpretable benchmark. The convergence of predictors across methods highlights the reliability of these findings and supports the integration of both acute physiological variables and chronic disease burden into early risk stratification frameworks.

Discussion

In this single-center, retrospective case–control study based on NTUH electronic health records, we developed and validated machine-learning models for predicting IHCA. To ensure comparability with the general inpatient population rather than a high-acuity subgroup at imminent risk of IHCA, random sampling was adopted for the control cohort. This strategy enabled us to construct a prediction model representative of routine hospitalized patients and to assess its performance in that context. Notably, malignancy was less common in the IHCA group-a paradoxical finding that may be explained by the higher prevalence of DNR orders among terminal cancer patients, thereby reducing their likelihood of unexpected IHCA.²⁰

Our findings highlight that combining traditional statistical approaches with modern ML methods provides complementary strengths in risk prediction. Logistic regression identified established clinical predictors, whereas ensemble models such as random forest and XGBoost achieved superior overall performance. These results underscore the value of integrating conventional regression with advanced ML in clinical prognostication.²¹

Feature importance analysis revealed complementary strengths. Logistic regression prioritized established predictors such as hemoglobin, pulse rate, ACS, heart failure, and platelet count, consistent with traditional cardiovascular frameworks.^5–7 In contrast, ML models consistently ranked BUN and corrected QT interval on ECG among the top variables, reflecting their ability to capture nonlinear relationships and complex interactions often overlooked by conventional approaches.^22,23 Together, these predictors, including BUN, corrected QT interval on ECG, hemoglobin, ACS, heart failure, platelet count, and inflammatory markers, illustrate the multifactorial nature of IHCA risk and underscore the value of integrating both chronic comorbidities and acute stressors into predictive models.^24,25

In this study, we adopted random sampling to construct the control group. This approach allowed us to better represent the heterogeneity of the general inpatient population and to identify the subgroup truly at risk of IHCA who might benefit from early intervention. In contrast, propensity score matching, while effective in reducing baseline imbalances, would restrict the analysis to patients already similar to the IHCA cohort based on predefined risk factors. Such restriction could limit generalizability and potentially overlook the broader at-risk population that our prediction models aim to capture.²⁶

Previous studies applying ML to IHCA prediction have reported AUCs of 0.80–0.93,^22,23,27 which are comparable to our results. One study demonstrated that gradient boosting outperformed logistic regression in emergency patients,²³ while another identified laboratory markers such as platelet count and serum sodium as powerful predictors,²⁷ aligning with our findings. Other investigations highlighted the predictive value of ECG-derived features such as corrected QT interval on ECG,^28–30 which was also confirmed in our analysis.

A conceptual strength of ML is its ability to move beyond binary “normal/abnormal” thresholds traditionally used in clinical medicine.^31–33 Logistic regression and conventional models depend on predefined cutoffs (eg, serum sodium <135 mmol/L) which may obscure risk gradients within reference ranges.³⁴ In contrast, ML derives optimal cut points directly from data. In our decision tree, BUN at 27 mg/dL emerged as a critical threshold for IHCA risk, despite lying near the conventional upper limit of normal. Similar data-driven thresholds were identified for hemoglobin (10 g/dL) and pulse rate (84 or 121 bpm). Such findings illustrate how ML can uncover hidden nonlinear risk profiles, as demonstrated in sepsis,^35,36 ACS,³⁷ and arrhythmia prediction.^27,30 For example, in Figure 3, the decision tree identified a diastolic blood pressure (DBP) threshold of 84 mmHg, which is not a commonly used clinical cut-off in daily practice. Nevertheless, prior studies have demonstrated that DBP is indeed an independent predictor of cardiac arrest, albeit with different threshold values.^38,39 This finding underscores the potential of ML models to uncover clinically relevant yet unconventional patterns that may be overlooked by traditional approaches. While such thresholds may not immediately translate into bedside decision rules, they highlight physiological parameters that warrant closer monitoring and further validation in prospective studies.

Beyond IHCA, ML models have been widely used for disease prediction across medicine. Decision trees are simple and transparent but often lack sensitivity in high-risk settings.³⁹ Random forest, by combining multiple trees, improves stability and has shown strong performance in predicting sepsis, ACS, and heart failure.⁴⁰ XGBoost, an advanced gradient boosting method, consistently outperforms other algorithms in structured healthcare datasets by capturing complex nonlinear relationships with high efficiency.⁴¹ Although less commonly used, MARS provides flexibility in modeling both linear and nonlinear effects. A previous study demonstrated its predictive value by developing a model for summed stress score in Taiwanese women with type 2 diabetes mellitus using the MARS approach.⁴²

Comparative studies confirm that ensemble methods, particularly random forest and XGBoost, provide the best overall accuracy and calibration, while decision trees and MARS contribute interpretability in selected scenarios.^40–42 Our findings echo prior evidence of XGBoost’s superiority and further support the robustness of ML models across diverse patient populations and healthcare systems. Importantly, when integrated into electronic health records, ML-based prediction tools could be embedded within hospital early warning systems to deliver real-time alerts and facilitate timely clinical intervention.¹⁴

A key challenge for implementing ML in clinical practice is interpretability, as advanced models often act as “black boxes” compared with the transparency of logistic regression.³² In addition, successful adoption requires seamless integration into electronic health record systems, with real-time outputs that are clinically actionable.⁴³ Overcoming these barriers will be crucial for translating predictive accuracy into meaningful patient outcomes.

We believe our study makes two main contributions. First, we systematically compared the performance of multiple machine learning models against traditional logistic regression, highlighting their relative strengths in predicting IHCA. Second, by applying multiple predictive tools, we were able to identify novel risk factors that are not typically captured by conventional approaches, and to establish an early warning framework that may help deliver intensive care to high-risk patients and thereby reduce mortality.

This study has several limitations. First, its retrospective, single-center design precludes causal inference and may limit generalizability. Second, we adopted random sampling rather than propensity score matching to ensure representativeness of the general inpatient population. This approach introduced baseline imbalances, but machine learning methods, with their ability to model multicollinearity and interactions, may have mitigated some of these differences. Third, only internal validation was performed; external, multicenter validation is needed to confirm robustness. Fourth, certain relevant variables (eg, echocardiography, Holter monitoring, imaging) were unavailable, which may influence risk assessment. Finally, as a pilot study, future research should incorporate multimodal data and prospective designs, ideally comparing model predictions with physicians’ real-time judgment, to establish clinical utility.

Conclusion

In this study, we directly compared logistic regression with multiple machine learning models for predicting in-hospital cardiac arrest. While logistic regression provided interpretability, advanced models-particularly XGBoost and random forest-achieved superior discrimination and calibration. Key predictors consistently included BUN, corrected QT interval, and hemoglobin. These results suggest that ML-based tools can enhance early risk stratification beyond conventional approaches, and their integration into hospital electronic health records and early warning systems may facilitate earlier recognition and timely intervention. Prospective multicenter validation will be essential to confirm these findings and determine their clinical impact.

Acknowledgments

The authors sincerely appreciate the data resources made available through the Integrated Medical Database of National Taiwan University Hospital, as well as the kind support offered by its staff. We are also indebted to the Artificial Intelligence Development Center at Fu Jen Catholic University, New Taipei City, Taiwan, for their valuable technical assistance.

This paper was previously uploaded to ResearchGate as a preprint [https://www.researchgate.net/publication/395063593_Comparative_Performance_of_Machine_Learning_Algorithms_and_Logistic_Regression_for_Predicting_In-Hospital_Cardiac_Arrest_Preprint]. It was initially submitted to JMIR Cardio but was formally withdrawn prior to its current submission.

Disclosure

The authors report no conflicts of interest in this work.

References

1. Liu C-T, Lai C-Y, Wang J-C, Chung C-H, Chien W-C, Tsai C-S. A population-based retrospective analysis of post-in-hospital cardiac arrest survival after modification of the chain of survival. J Emerg Med. 2020;59(2):246–253. doi:10.1016/j.jemermed.2020.04.045

2. Peberdy MA, Kaye W, Ornato JP, et al. Cardiopulmonary resuscitation of adults in the hospital: a report of 14720 cardiac arrests from the national registry of cardiopulmonary resuscitation. Resuscitation. 2003;58(3):297–308. doi:10.1016/s0300-9572(03)00215-6

3. Merchant RM, Yang L, Becker LB, et al. Incidence of treated cardiac arrest in hospitalized patients in the United States. Crit Care Med. 2011;39(11):2401–2406. doi:10.1097/CCM.0b013e3182257459

4. Wang CH, Tay J, Wu CY, et al. External validation and comparison of statistical and machine learning-based models in predicting outcomes following out-of-hospital cardiac arrest: a multicenter retrospective analysis. J Am Heart Assoc. 2024;13(20):e037088. doi:10.1161/JAHA.124.037088

5. Girotra S, Nallamothu BK, Spertus JA, et al. Trends in survival after in-hospital cardiac arrest. N Engl J Med. 2012;367(20):1912–1920. doi:10.1056/NEJMoa1109148

6. Nolan JP, Soar J, Smith GB, et al. National cardiac arrest audit. Incidence and outcome of in-hospital cardiac arrest in the United Kingdom national cardiac arrest audit. Resuscitation. 2014;85(8):987–992. doi:10.1016/j.resuscitation.2014.04.002

7. Guan G, Lee CMY, Begg S, Crombie A, Mnatzaganian G. The use of early warning system scores in prehospital and emergency department settings to predict clinical deterioration: a systematic review and meta-analysis. PLoS One. 2022;17(3):e0265559. doi:10.1371/journal.pone.0265559

8. Smith GB, Prytherch DR, Meredith P, Schmidt PE, Featherstone PI. The ability of the National Early Warning Score (NEWS) to discriminate patients at risk of early cardiac arrest, unanticipated intensive care unit admission, and death. Resuscitation. 2013;84(4):465–470. doi:10.1016/j.resuscitation.2012.12.016

9. Smith ME, Chiovaro JC, O’Neil M, et al. Early warning system scores for clinical deterioration in hospitalized patients: a systematic review. Ann Am Thorac Soc. 2014;11(9):1454–1465. doi:10.1513/AnnalsATS.201403-102OC

10. Churpek MM, Yuen TC, Edelson DP. Risk stratification of hospitalized patients on the wards. Chest. 2013;143(6):1758–1765. doi:10.1378/chest.12-1605

11. Badriyah T, Briggs JS, Meredith P, et al. Decision-tree early warning score (DTEWS) validates the design of the National Early Warning Score (NEWS). Resuscitation. 2014;85(3):418–423. doi:10.1016/j.resuscitation.2013.12.011

12. Hogan H, Hutchings A, Wulff J, et al. Interventions to Reduce Mortality from in-Hospital Cardiac Arrest: A Mixed-Methods Study. Southampton (UK): NIHR Journals Library; January 2019.

13. Shafiq M, Mazzotti DR, Gibson C. Risk stratification of patients who present with chest pain and have normal troponins using a machine learning model. World J Cardiol. 2022;14(11):565–575. doi:10.4330/wjc.v14.i11.565

14. Rajkomar A, Oren E, Chen K, et al. Scalable and accurate deep learning with electronic health records. NPJ Digit Med. 2018;1:18. doi:10.1038/s41746-018-0029-1

15. Weng SF, Reps J, Kai J, Garibaldi JM, Qureshi N. Can machine learning improve cardiovascular risk prediction using routine clinical data? PLoS One. 2017;12(4):e0174944. doi:10.1371/journal.pone.0174944

16. Chen CT, Chiu PC, Tang CY, et al. Prognostic factors for survival outcome after in-hospital cardiac arrest: an observational study of the oriental population in Taiwan. J Chin Med Assoc. 2016;79(1):11–16. doi:10.1016/j.jcma.2015.07.011

17. Andersen LW, Holmberg MJ, Berg KM, Donnino MW, Granfeldt A. In-hospital cardiac arrest: a review. JAMA. 2019;321(12):1200–1210. doi:10.1001/jama.2019.1696

18. Fernando SM, Tran A, Cheng W, et al. Pre-arrest and intra-arrest prognostic factors associated with survival after in-hospital cardiac arrest: systematic review and meta-analysis. BMJ. 2019:367:l6373. doi:10.1136/bmj.l6373

19. Mitsunaga T, Hasegawa I, Uzura M, et al. Comparison of the National Early Warning Score (NEWS) and the Modified Early Warning Score (MEWS) for predicting admission and in-hospital mortality in elderly patients in the prehospital setting and in the emergency department. PeerJ. 2019;7(e6947). doi:10.7717/peerj.6947

20. Giza DE, Graham J, Donisan T, et al. Impact of cardiopulmonary resuscitation on survival in cancer patients: do not resuscitate before or after CPR? JACC CardioOncol. 2020;2(2):359–362. doi:10.1016/j.jaccao.2020.03.003

21. Holmstrom L, Bednarski B, Chugh H, et al. Artificial intelligence model predicts sudden cardiac arrest manifesting with pulseless electric activity versus ventricular fibrillation. Circ Arrhythm Electrophysiol. 2024;17(2):e012338. doi:10.1161/CIRCEP.123.012338

22. Kwon JM, Kim KH, Jeon KH, Lee SY, Park J, Oh BH. Artificial intelligence algorithm for predicting cardiac arrest using electrocardiography. Scand J Trauma Resusc Emerg Med. 2020;28(1):98. doi:10.1186/s13049-020-00791-0

23. Lu TC, Wang CH, Chou FY, et al. Machine learning to predict in-hospital cardiac arrest from patients presenting to the emergency department. Intern Emerg Med. 2023;18(2):595–605. doi:10.1007/s11739-022-03143-1

24. Weng SF, Vaz L, Qureshi N, Kai J. Prediction of premature all-cause mortality: a prospective general population cohort study comparing machine learning and standard epidemiological approaches. PLoS One. 2019;14(3):e0214365. doi:10.1371/journal.pone.0214365

25. Li H, Wu TT, Yang DL, et al. Decision tree model for predicting in-hospital cardiac arrest among patients admitted with acute coronary syndrome. Clin Cardiol. 2019;42(11):1087–1093. doi:10.1002/clc.23255

26. Stürmer T, Wyss R, Glynn RJ, Brookhart MA. Propensity scores for confounder adjustment when assessing the effects of medical interventions using nonexperimental study designs. J Intern Med. 2014;275(6):570–580. doi:10.1111/joim.12197

27. Ding X, Wang Y, Ma W, et al. Development of early prediction model of in-hospital cardiac arrest based on laboratory parameters. Biomed Eng Online. 2023;22(1):116. doi:10.1186/s12938-023-01178-9

28. Do DH, Kuo A, Lee ES, et al. Usefulness of trends in continuous electrocardiographic telemetry monitoring to predict in-hospital cardiac arrest. Am J Cardiol. 2019;124(7):1149–1158. doi:10.1016/j.amjcard.2019.06.032

29. Straus SM, Kors JA, De Bruin ML, et al. Prolonged QTc interval and risk of sudden cardiac death in a population of older adults. J Am Coll Cardiol. 2006;47(2):362–367. doi:10.1016/j.jacc.2005.08.067

30. Al-Khatib SM, LaPointe NM, Kramer JM, Califf RM. What clinicians should know about the QT interval. JAMA. 2003;289(16):2120–2127. doi:10.1001/jama.289.16.2120

31. Matsushita K, Ballew SH, Wang AY, et al. Epidemiology and risk of cardiovascular disease in populations with chronic kidney disease. Nat Rev Nephrol. 2022;18(11):696–707. doi:10.1038/s41581-022-00616-6

32. Rajkomar A, Dean J, Kohane I. Machine learning in medicine. N Engl J Med. 2019;380(14):1347–1358. doi:10.1056/NEJMra1814259

33. Misra D, Avula V, Wolk DM, et al. Early detection of septic shock onset using interpretable machine learners. J Clin Med. 2021;10(2):301. doi:10.3390/jcm10020301

34. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206–215. doi:10.1038/s42256-019-0048-x

35. Jin D, Jin S, Liu B, et al. Association between serum sodium and in-hospital mortality among critically ill patients with spontaneous subarachnoid hemorrhage. Front Neurol. 2022;13:1025808. doi:10.3389/fneur.2022.1025808

36. Seymour CW, Kennedy JN, Wang S, et al. Derivation, validation, and potential treatment implications of novel clinical phenotypes for sepsis. JAMA. 2019;321(20):2003–2017. doi:10.1001/jama.2019.5791

37. VanHouten JP, Starmer JM, Lorenzi NM, Maron DJ, Lasko TA. Machine learning for risk prediction of acute coronary syndrome. AMIA Annu Symp Proc. 2014;2014:1940–1949.

38. Aziz S, Barratt J, Starr Z, et al. The association between intra-arrest arterial blood pressure and return of spontaneous circulation in out-of-hospital cardiac arrest. Resuscitation. 2024;205:110426. doi:10.1016/j.resuscitation.2024.110426

39. Zelic I, Kononenko I, Lavrac N, Vuga V. Induction of decision trees and Bayesian classification applied to diagnosis of sport injuries. J Med Syst. 1997;21(6):429–444. doi:10.1023/A:1022880431298

40. Goldstein BA, Navar AM, Carter RE. Moving beyond regression techniques in cardiovascular risk prediction: applying machine learning to address analytic challenges. Eur Heart J. 2017;38(23):1805–1814. doi:10.1093/eurheartj/ehw302

41. Wang Z, Gu Y, Huang L, et al. Construction of machine learning diagnostic models for cardiovascular pan-disease based on blood routine and biochemical detection data. Cardiovasc Diabetol. 2024;23(1):351. doi:10.1186/s12933-024-02439-0

42. Yuan CH, Lee PC, Wu ST, Yang CC, Chu TW, Yeih DF. Using multivariate adaptive regression splines to estimate summed stress score on myocardial perfusion scintigraphy in Chinese women with type 2 diabetes: a comparative study with multiple linear regression. Diagnostics. 2025;15(17):2270. doi:10.3390/diagnostics15172270

43. Hofer IS, Burns M, Kendale S, Wanderer JP. Realistically integrating machine learning into clinical practice: a road map of opportunities, challenges, and a potential future. Anesth Analg. 2020;130(5):1115–1118. doi:10.1213/ANE.0000000000004575

Introduction

Natural childbirth is widely recognized as a relatively safe delivery method for mothers, offering faster postpartum recovery and effectively avoiding the short- and long-term complications associated with cesarean sections.^1–3 However, the labor process for natural delivery is often prolonged, particularly for first-time mothers, and it is difficult to avoid pain caused by uterine contractions during labor. This pain not only impacts the delivery process but may also pose certain risks to maternal and neonatal safety.^4–6 Labor pain, one of the primary physiological challenges faced by mothers during childbirth, primarily stems from uterine contractions, cervical dilation, and pressure on the birth canal. Such pain can trigger significant psychological stress in mothers, potentially reducing the efficiency of contractions, prolonging labor duration, and consequently increasing the likelihood of cesarean delivery and other complications.^7–9 Although traditional supine delivery facilitates medical monitoring, it has been associated with potential disadvantages, including increased reports of maternal discomfort, compromised hemodynamics (eg, supine hypotensive syndrome), and potential narrowing of the pelvic outlet compared to upright positions, potentially exacerbating maternal discomfort and hindering fetal descent.^10–12

In recent years, non-pharmacological interventions have gained increasing attention in obstetric research and clinical practice as part of a broader movement towards humanized childbirth care.¹³ Studies have shown that evidence-based, woman-centered labor care not only facilitates smoother delivery but also effectively reduces the risk of adverse outcomes.^14,15 Among these interventions, free positioning and mindful relaxation techniques have garnered significant attention for their role in supporting the delivery process. Free positioning during labor transcends the limitations of traditional supine delivery, allowing mothers to choose positions such as standing, squatting, kneeling, or lying on their sides according to their needs and comfort. This autonomy in movement is thought to optimize the pelvic angle (potentially increasing the anteroposterior diameter), facilitate the descent and rotation of the fetus by utilizing gravity, and may reduce soft tissue resistance, thereby potentially reducing delivery difficulties and associated pain.^16–18 However, widespread implementation of free positioning faces challenges, including staff training requirements, resource constraints (eg, availability of birthing aids like balls or stools), and adherence to conventional protocols in some settings.^14,19 Meanwhile, mindful relaxation techniques (MRTs), often rooted in Mindfulness-Based Stress Reduction (MBSR) principles adapted for childbirth, involve deliberate attention regulation, non-judgmental awareness of present-moment sensations (including pain), breathing regulation, progressive muscle relaxation, and positive emotional guidance.^20,21 The core aim of MRTs is to help mothers modulate their psychological and physiological stress responses, thereby reducing the affective and cognitive components of pain perception, alleviating anxiety, and enhancing their overall sense of control and childbirth experience.^22,23 Despite documented benefits, the integration of structured MRTs into routine antenatal education and intrapartum care remains variable and is not yet standard practice globally.^24,25

Although previous studies have separately explored the roles of free positioning and mindful relaxation techniques in childbirth, rigorous research evaluating their combined application in natural delivery remains notably scarce. Systematic reviews highlight the independent benefits but lack robust evidence on synergistic effects.^26,27 Considering the complementary physiological and psychological effects of these two approaches, a synergistic interaction is theoretically plausible: free positioning primarily addresses biomechanical factors and physiological pain pathways (eg, optimizing fetal position and reducing soft tissue strain), while MRTs target the psychological and neurocognitive aspects of pain processing (eg, reducing fear-tension-pain cycles and enhancing pain coping mechanisms).^28,29 This combined approach may lead to greater overall pain reduction and labor efficiency than either intervention alone. Therefore, this study aims to examine the combined intervention’s efficacy in alleviating labor pain and shortening labor duration, thereby providing a basis for further optimization of obstetric care strategies. This investigation is particularly relevant within our institutional context in China, where there is a growing national emphasis on promoting natural childbirth and enhancing maternal satisfaction within the healthcare system,³⁰ yet barriers to implementing comprehensive non-pharmacological approaches persist.

Subjects and Methods

Study Subjects

Study Design and Ethical Considerations

This retrospective cohort study included pregnant women who delivered at Wuxi Second People’s Hospital between August 2023 and October 2024. The study was approved by the Ethics Committee of Wuxi Second People’s Hospital (Approval No.: 2022–081) and conducted in accordance with the 1964 Helsinki Declaration and its later amendments. Due to the retrospective nature of the study, the requirement for informed consent was waived by the IRB. Data collection employed an “opt-out” method, consistent with national regulations and ethical standards for minimal-risk retrospective research using anonymized data. Potential participants were informed about the study via hospital noticeboards and electronic platforms; those declining participation within two weeks were excluded. No objections were registered.

All collected data were strictly confidential and anonymized before analysis to protect participants’ privacy. Access to the data was limited to authorized research personnel only, and all data handling complied with institutional and national data protection regulations to ensure data security throughout the study.

Participants

After screening electronic medical records against predefined criteria, a total of 120 eligible participants were identified and included in the analysis. Participants were categorized into two groups based on the documented care received during delivery: the research group (n = 60), who received the combined intervention of free positioning and mindful relaxation techniques, and the Control Group (n = 60), who received routine delivery care.

Inclusion and Exclusion Criteria

Inclusion Criteria

Participants were required to meet the following criteria:

Singleton pregnancy at full-term gestation;

No absolute indications for cesarean delivery and the ability to undergo natural childbirth;

Availability of complete and traceable clinical records;

Normal communication abilities and clear consciousness.

Exclusion Criteria

Participants were excluded if they met any of the following conditions:

Presence of pregnancy complications such as hypertensive disorders of pregnancy or gestational diabetes;

High-risk pregnancies, including advanced maternal age or multiple pregnancies;

Abnormal pelvic structure or birth canal deformities;

Coagulation disorders or cognitive impairments;

Concurrent cardiac, pulmonary, or other major organ dysfunction.

Interventions

Control Group: Routine Delivery Care

Participants received standard obstetric care per hospital protocol. This included continuous fetal heart rate monitoring and assessment of cervical dilation. Upon reaching active labor (cervix dilated ≥3 cm), participants were transferred to the delivery room. Midwives provided continuous presence and support. Participants were typically guided to adopt supine or semi-recumbent positions as labor progressed and were instructed on breathing and pushing techniques during the second stage. Psychological support included verbal encouragement and physical reassurance (eg, hand-holding, gentle stroking).

Research Group: Combined Intervention (Free Positioning + Mindful Relaxation)

Midwives delivering this intervention had completed a standardized 40-hour training program covering: 1) Theory and evidence for free positioning and mindful relaxation; 2) Practical demonstration and supervised practice of all positions and mindfulness techniques; 3) Communication skills for guiding women; and 4) Documentation requirements. Training was based on WHO recommendations²² and FIGO guidelines.³¹

Free Positioning: During the active phase of labor (cervical dilation ≥3 cm), midwives introduced and demonstrated a variety of labor positions to the participants. Women were encouraged to move freely and select positions based on their comfort and individual preference, with midwives providing support and assistance for any necessary position changes. Commonly adopted positions included: recumbent positions (semi-recumbent, right lateral, and left lateral), standing positions (standing upright beside the bed while holding onto support), sitting positions (seated on the delivery bed or a low stool with hands supporting and body slightly leaning back), squatting positions (feet apart, supported by holding onto a chair or bed edge, often with midwife assistance), and kneeling positions (kneeling on a soft mat with knees apart and leaning forward on a pillow or blanket). There were no fixed time requirements for any specific posture; changes in position occurred as frequently as desired by the woman or in response to discomfort, allowing for a flexible, real-world application of the intervention.

Mindful Relaxation Techniques (MRTs): Upon admission to the delivery room (with cervical dilation ≥3 cm), guided mindful relaxation sessions were initiated and offered continuously throughout labor, particularly during contractions. Each session typically lasted 10–20 minutes and was flexibly adjusted to align with the pattern of uterine contractions and the participant’s level of concentration. While maintaining a self-selected labor position, participants were guided by midwives through a structured mindfulness protocol. This included sensory focus, wherein women were provided with a real object (such as an apple or orange) and instructed to observe its color, shape, and texture, touch its surface, and appreciate its aroma. This was followed by internalization, where they were guided to close their eyes, mentally recall the object’s characteristics, and focus attention while clearing the mind of distractions. A calm ambiance was created using continuous instrumental music played at a moderate volume (40–60 dB). During guided imagery, participants were encouraged to use their imagination to mentally explore the object in greater detail, promoting immersive concentration. In the breath regulation phase, midwives guided a transition from shallow, rapid breathing to deep, steady diaphragmatic breathing, synchronized with silent counting from one to ten and back. Finally, in the reflection stage, participants were instructed to open their eyes, then close them again to recall and verbalize their experience. This protocol, adapted from Mindfulness-Based Childbirth and Parenting (MBCP) principles,³² aimed to alleviate maternal stress, enhance emotional and physical relaxation, and improve women’s coping with labor pain.

Observation Indicators

In this retrospective cohort study, all observation indicators were extracted from standardized medical and nursing records. To reduce bias arising from inconsistent data collection, especially for subjective measures (eg, pain or psychological states), only data recorded by trained staff using standardized tools during the clinical process were included. Records with missing or retrospectively added assessments were excluded to enhance data reliability. The specific indicators evaluated were as follows:

Duration of Labor

Labor duration was segmented into the first, second, and third stages, with total labor time subsequently calculated. These data were obtained from partogram records maintained by midwives in real time during delivery. Comparative analysis was conducted between the intervention and control groups to evaluate differences in labor progression.

Pain Levels

Pain intensity during labor was evaluated using a multidimensional framework based on the Verbal Rating Scale (VRS) recommended by the World Health Organization in 1980. This included three components: (1) the VRS, which classifies pain into four levels—0 (no pain), 1 (mild pain that does not interfere with sleep or daily life), 2 (moderate pain requiring non-narcotic interventions), and 3 (severe pain requiring narcotic analgesia and often accompanied by autonomic symptoms); (2) the Pain Rating Index (PRI), consisting of 15 descriptors, each rated from 0 to 3 (no, mild, moderate, or severe pain), with higher scores indicating greater overall pain perception; and (3) the Visual Analogue Scale (VAS), a 10-cm horizontal ruler labeled from 0 (no pain) to 10 (worst imaginable pain), where participants marked the intensity of their pain. Only assessments recorded by midwives during labor were included in the analysis.

Perineal Tears

The degree of perineal trauma was classified based on clinical records according to standard obstetric grading criteria. These included: intact perineum (no laceration), Grade I (superficial tears of the vaginal or perineal mucosa), Grade II (tears involving the perineal muscle layer and/or posterior vaginal wall), Grade III (tears extending to the external anal sphincter), and Grade IV (full-thickness tears involving the rectal mucosa). Grading was performed by attending midwives or obstetricians immediately postpartum and documented in delivery notes.

Sense of Labor Control

Maternal sense of control during childbirth was measured using the Labor Agentry Scale (LAS), which comprises 29 items rated on a 7-point Likert scale, yielding a total score between 29 and 203. Higher scores represent greater perceived control and agency during labor. In this retrospective analysis, only LAS questionnaires completed within two hours postpartum and documented in full by trained staff were considered valid.

Negative Emotions

Psychological states, specifically anxiety and depression, were assessed using the Self-Rating Anxiety Scale (SAS) and the Self-Rating Depression Scale (SDS), respectively. Each scale includes 20 items rated on a 4-point scale, with a total score of 80. Higher scores indicate more severe emotional distress. In order to ensure consistency, only data collected at standardized time points (within one week before intervention and within 30 minutes postpartum) and documented under supervision were included in the analysis.

Pregnancy Outcomes

Pregnancy outcomes were assessed by recording postpartum hemorrhage volume within two hours of delivery and evaluating neonatal health using the Apgar scoring system. The Apgar score includes five parameters—skin color, heart rate, respiratory effort, muscle tone, and reflex response—each rated from 0 to 2, with a maximum total of 10. Scores were categorized as 10 (optimal condition), 7–9 (mild concerns), or <7 (requires immediate intervention). These scores were extracted from standardized neonatal assessment forms filled by neonatal nurses or pediatricians at 1 and 5 minutes after birth.

Data Analysis

A priori sample size calculation was performed using G*Power 3.1.²⁵ Based on pilot data (unpublished) and previous studies,^6,18 a medium effect size (Cohen’s d = 0.65) for the primary outcome (pain VAS score) was assumed. To achieve 80% power (α=0.05, two-tailed t-test), 60 participants per group were required. This justified the final sample size of n=120 (60 per group).

GraphPad Prism 8 was used for graphical presentation. SPSS 26.0 was used for statistical analysis. Quantitative data were assessed for normality using the Shapiro–Wilk test and visual inspection of Q-Q plots. Normally distributed data are presented as Mean ± Standard Deviation (SD) and compared using Independent Samples t-tests. Non-normally distributed data are presented as Median (Interquartile Range, IQR) and compared using Mann–Whitney U-tests. Qualitative data are presented as Number (Percentage, %) and compared using Chi-square (χ²) or Fisher’s exact test, as appropriate. Effect sizes are reported (Cohen’s d for t-tests, Cramer’s V for χ²). A two-tailed P value < 0.05 was considered statistically significant. No adjustments for multiple comparisons were made for secondary outcomes, consistent with exploratory analysis in this retrospective design; findings should be interpreted accordingly.

Results

Baseline Data

The control group included 60 participants, aged 25–35 years (Mean ± SD: 28.44 ± 2.63 years); gestational age 37–42 weeks (39.02 ± 0.94 weeks); body weight 50–78 kg (63.15 ± 6.22 kg).

The research group included 60 participants, aged 25–35 years (28.96 ± 2.58 years); gestational age 37–42 weeks (39.07 ± 1.02 weeks); body weight 50–78 kg (63.74 ± 6.18 kg).

No significant differences existed in baseline characteristics (P > 0.05, Table 1), confirming comparability.

Table 1 Comparison of Baseline Data Between the Two Groups (Mean ± SD)

Labor Duration

As shown in Figure 1, the research group exhibited significantly shorter labor durations across all three stages compared to the control group. Specifically, the mean duration of the first stage of labor was markedly reduced in the research group (240.69 ± 25.96 minutes) compared to the control group (362.47 ± 31.94 minutes), with a large effect size (Cohen’s d = 4.22, 95% CI: 112.15–131.41, p < 0.001). The second stage of labor was also significantly shorter in the research group (42.32 ± 10.69 minutes) versus the control group (52.14 ± 12.23 minutes), with a moderate to large effect size (Cohen’s d = 0.87, 95% CI: 6.29–13.35, p < 0.001). For the third stage, the research group showed a mean duration of 5.11 ± 2.56 minutes, significantly less than the control group’s 8.74 ± 2.85 minutes (Cohen’s d = 1.35, 95% CI: 2.78–4.48, p < 0.001). Overall, the total labor duration was significantly reduced by approximately 135 minutes in the research group (291.23 ± 28.65 minutes) compared to the control group (426.56 ± 40.69 minutes), with a large effect size (Cohen’s d = 3.86, 95% CI: 121.30–149.36, p < 0.001).

Figure 1 Comparison of Labor Duration Between the Two Groups. Note: *Indicates P < 0.05 compared to the control group.

Pain Levels

As shown in Figure 2, pain levels following the intervention were significantly lower in the research group across all assessment dimensions. On the Visual Analog Scale (VAS), the research group reported a mean score of 4.32 ± 1.03, significantly lower than the control group’s 5.23 ± 1.24 (Cohen’s d = 0.80, 95% CI: 0.65–1.17, p < 0.001), indicating a moderate to large effect size. Similarly, the Pain Rating Index (PRI) scores were reduced in the research group (2.49 ± 0.45) compared to the control group (2.94 ± 0.78), with a moderate effect size (Cohen’s d = 0.71, 95% CI: 0.28–0.62, p < 0.001). In terms of Present Pain Intensity (PPI), the research group scored 25.45 ± 2.14, significantly lower than the control group’s 28.77 ± 2.56 (Cohen’s d = 1.42, 95% CI: 2.67–3.97, p < 0.001), reflecting a large effect size. These findings indicate that the combined intervention of free positioning and mindful relaxation techniques was effective in significantly alleviating both the sensory and cognitive-affective components of labor pain.

Figure 2 Comparison of VRS Scores Between the Two Groups. Note: *Indicates a significant difference between the two groups, P<0.05.

Perineal Laceration

As illustrated in Figure 3, the distribution of perineal outcomes differed significantly between the two groups (P < 0.05). In the research group, 35.00% (21/60) of participants had an intact perineum, 41.67% (25/60) experienced Grade I lacerations, 21.67% (13/60) had Grade II lacerations, and 1.67% (1/60) sustained Grade III lacerations. In comparison, the control group showed 23.33% (14/60) with an intact perineum, 30.00% (18/60) with Grade I lacerations, 35.00% (21/60) with Grade II lacerations, and 11.67% (7/60) with Grade III lacerations. These findings indicate a higher proportion of intact perineum and first-degree lacerations but a lower incidence of more severe (Grade II and III) tears in the research group, suggesting a protective effect of the intervention on perineal outcomes.

Figure 3 Comparison of Perineal Laceration Degrees Between the Two Groups. Note: *Indicates a significant difference between the two groups, P<0.05.

Labor Control

The analysis of Labor Agentry Scale (LAS) scores revealed a statistically significant difference between the two groups. Participants in the research group reported a higher sense of control during labor, with a mean LAS score of 152.41 ± 8.11, compared to 144.22 ± 9.11 in the control group (d = 0.95; 95% CI: 5.42–10.96; P < 0.001). This suggests that the combined intervention notably enhanced participants’ perceived autonomy and empowerment throughout the birthing process. See Table 2.

Table 2 Comparison of LAs Scores Between the Two Groups (Mean ± SD)

Negative Emotions

The analysis of post-intervention psychological outcomes indicated that participants in the research group experienced significantly lower levels of anxiety and depression compared to those in the control group. Specifically, the Self-Rating Anxiety Scale (SAS) scores were 46.23 ± 2.35 in the research group versus 55.98 ± 2.47 in the control group (t = 22.152, P < 0.001; d = 4.10; 95% CI: 8.99–10.23). Similarly, Self-Rating Depression Scale (SDS) scores were 45.22 ± 1.73 in the research group and 54.56 ± 2.14 in the control group (t = 26.291, P < 0.001; d = 4.71; 95% CI: 8.63–10.06). These findings suggest that the combined intervention was highly effective in reducing negative emotional responses during labor. See Table 3.

Table 3 Comparison of Anxiety and Depression Scores Between the Two Groups ()

Pregnancy Outcomes

Postpartum outcomes revealed significant differences between the two groups. The mean volume of bleeding within 2 hours postpartum was markedly lower in the research group (155.89 ± 22.21 mL) compared to the control group (204.58 ± 30.88 mL), indicating improved hemostatic outcomes (d = 1.79; 95% CI: 40.65–57.59; P < 0.001). Apgar scores were slightly higher in the research group (8.44 ± 1.56) compared to the control group (8.05 ± 1.11); however, the difference was not statistically significant (d = 0.28; 95% CI: 0.13–0.73; P = 0.145), indicating comparable neonatal conditions between groups. See Figure 4.

Figure 4 Comparison of Postpartum 2-Hour Blood Loss and Neonatal Apgar Scores Between the Two Groups. Note: *Indicates a significant difference between the two groups, P<0.05.

Discussion

This study aimed to evaluate the effects of free positioning combined with mindfulness relaxation techniques on labor pain relief and labor duration reduction. The results showed that this combined intervention significantly shortened the duration of each stage of labor, alleviated labor pain, improved maternal sense of control and emotional status during labor, and reduced postpartum blood loss, demonstrating favorable clinical outcomes. These findings are consistent with multiple previous studies, further validating the value of this integrated intervention model in promoting natural childbirth.

Firstly, the significant shortening of labor duration is one of the core findings of this study. The intervention group’s average duration of the first stage of labor was markedly shorter than that of the control group (240.69 minutes vs 362.47 minutes), with an overall labor time reduction of approximately 135 minutes and a very large effect size (Cohen’s d = 3.86). This result aligns with the conclusions of Mansfield et al’s systematic review on free positioning facilitating labor progress, which indicated that freedom of movement and position changes could reduce the risk of prolonged cervical dilation and extended second stage of labor.³³ The mechanism may be related to multiple factors: free positioning allows the mother to utilize gravity to assist fetal descent, relieve pressure on the birth canal during uterine contractions, facilitate better alignment of the fetal head with the birth canal curve, and reduce birth canal resistance.³⁴ Moreover, alternating between different positions can improve pelvic morphology and blood circulation, helping relax the pelvic floor muscles and thus accelerating labor.³⁵

In addition, the effect of mindfulness relaxation in alleviating labor pain and improving emotional state may also partly be attributed to the potential regulation of the neuroendocrine system. Although this study did not directly measure neuroendocrine-related indicators, existing literature shows that mindfulness interventions can reduce sympathetic nervous system activity, enhance parasympathetic tone, and regulate the hypothalamic-pituitary-adrenal (HPA) axis reactivity, thereby decreasing the secretion of stress hormones such as cortisol and norepinephrine.^36–38 This neuroendocrine balance adjustment helps lower pain sensitivity, relieve anxiety and tension, increase pain threshold, and enhance maternal sense of control and adaptability during labor. In this study, women in the intervention group performed better in subjective pain scores (VAS, PPI, etc)., anxiety and depression scales (SAS/SDS), and labor control scores (LAS), indirectly supporting the plausibility of this mechanism. This dual physiological–psychological pathway suggests that the combined intervention of free positioning and mindfulness relaxation not only improves clinical outcomes but may also have favorable neuroregulatory and mind-body synergistic effects, warranting further exploration at the physiological mechanism level.

Simultaneously, the intervention group showed a significantly enhanced sense of labor control (LAS scores significantly higher than control), indicating that this intervention model improved maternal autonomy and self-efficacy. A high sense of control during labor is closely related to reduced labor anxiety, decreased frequency of obstetric interventions, and promotion of vaginal delivery.³⁹ This is because free positioning and mindfulness relaxation allow mothers to actively choose comfortable postures and psychological adjustment methods, increasing their mastery over the delivery process and reducing fear and helplessness.⁴⁰ This result is also supported by studies by Li and Guo, who reported that psychological support and self-regulation strategies during labor significantly improve maternal delivery satisfaction and psychological health.^41,42

This study also observed a significantly lower incidence of severe perineal lacerations (grade II and above) and a notable reduction in postpartum blood loss in the intervention group. Free positioning (eg, semi-sitting, lateral, squatting) can reduce perineal tension, promote natural soft tissue expansion, and decrease the risk of mechanical injury.⁴³ Mindfulness relaxation may contribute by lowering stress responses, improving vascular tone and tissue perfusion, thereby facilitating local repair and hemostasis.⁴⁴ These findings are consistent with Hughes’ review on perineal protection strategies, which emphasizes the importance of posture adjustment and emotional interventions in reducing perineal trauma and postpartum hemorrhage.⁴⁵

In summary, this study confirms that free positioning combined with mindfulness relaxation techniques significantly promotes labor pain relief, shortens labor duration, reduces perineal trauma, and improves maternal psychological state. The mechanisms involve physiological and mechanical optimization (such as fetal descent and pelvic morphology changes), neuroendocrine regulation (reduced sympathetic excitation and enhanced analgesia), and psychological-behavioral enhancement of labor control. The synergy of these mechanisms improves the overall childbirth experience and facilitates smooth natural delivery.

Limitations

Although the results show that free positioning combined with mindfulness relaxation techniques have positive effects on labor pain relief, labor duration reduction, emotional improvement, and childbirth experience enhancement, several limitations should be fully considered when interpreting the findings. First, this study is a retrospective cohort design and is limited by the completeness and accuracy of existing medical records, which may introduce information and recall bias. Since randomization and blinding were not performed, there may be uncontrolled baseline differences and selection bias between the intervention and control groups, affecting the rigor of causal inference. Moreover, potential confounders such as maternal social support, fear of childbirth, experience level of birth attendants, and fetal position were not fully controlled, which could partially interfere with the assessment of intervention effects.

Second, the sample size is relatively small and drawn from a single center, which may affect the representativeness and generalizability of the results. Future studies should validate these findings through multicenter, large-sample, prospective randomized controlled trials to strengthen external validity and causal inference. Additionally, as this study did not include single free positioning or single mindfulness intervention groups, it cannot clearly evaluate the independent contribution of each component nor conclude whether true synergistic effects exist. Therefore, the term “combined intervention effect” should be used cautiously to indicate superior effects compared with routine care but not to prove interaction between interventions.

Finally, although literature suggests that mindfulness interventions may exert effects via neuroendocrine regulation—such as reducing sympathetic activity, modulating the HPA axis, and decreasing stress hormone secretion^36–38—this study did not directly measure neuroendocrine markers, so these remain hypothetical explanations that need further mechanistic research to validate their biological basis.

Conclusion

The results of this study preliminarily suggest that free positioning combined with mindfulness relaxation interventions may help shorten labor duration, relieve labor pain, improve emotional state, and enhance maternal sense of control during childbirth, thereby promoting smooth natural delivery. This non-pharmacological, low-cost intervention model has certain potential for clinical promotion, especially in resource-limited or humanized childbirth-focused settings.

However, given the retrospective design, lack of randomization, blinding, and long-term follow-up, causal interpretations should be cautious. Additionally, the synergistic effects of the combined intervention cannot be separated to clarify individual components’ independent effects. Future research should employ more rigorous prospective randomized controlled trials or factorial design studies to further elucidate intervention mechanisms and explore applicability and sustainability across different populations and labor stages.

In summary, the current findings provide valuable preliminary evidence supporting non-pharmacological interventions to promote natural childbirth, warranting further exploration and validation in higher-quality studies to assess feasibility and effectiveness for broader application.

Disclosure

The authors report no conflicts of interest in this work.

References

1. Michalska A, Niechcial K, Niechcial R, et al. Natural childbirth and cesarean section – descriptive analysis of queries in Google search engine. Ginekol Pol. 2024;95(7):565–572. doi:10.5603/gpl.97654

2. Sharma S, Dhakal I. Cesarean vs Vaginal Delivery: an Institutional Experience. J Nepal Med Assoc. 2018;56(209):535–539. doi:10.31729/jnma.3467

3. Cosans C. The meaning of natural childbirth. Perspect Biol Med. 2004;47(2):266–272. doi:10.1353/pbm.2004.0022

4. Buxton IL, Crow W, Mathew SO. Regulation of uterine contraction: mechanisms in preterm labor. AACN Clin Issues. 2000;11(2):271–282. doi:10.1097/00044067-200005000-00010

5. Palomar Morales M, Hicks JJ. [Uterine contraction. Mechanism, regulation and pharmacology]. Ginecol Obstet Mex. 1990;58:303–309.

6. Huang B, Lu Y, Zhang Y, et al. Application of natural shoulder delivery combined with free position delivery in maternal delivery. Am J Transl Res. 2021;13(12):14168–14175.

7. Guo H, Que M, Shen J, et al. Effect of music therapy combined with free position delivery on labor pain and birth outcomes. Appl Bionics Biomech. 2022;2022:8963656. doi:10.1155/2022/8963656

8. Hanson L, Verklan T, Bakewell-Sachs S. Pertinent to intrapartum nursing. J Perinat Neonatal Nurs. 2014;28(2):89–90. doi:10.1097/JPN.0000000000000036

9. Anderson G, Zega M, D’Agostino F, et al. Meta-synthesis of the needs of women cared for by midwives during childbirth in hospitals. J Obstet Gynecol Neonatal Nurs. 2021;50(1):6–19. doi:10.1016/j.jogn.2020.10.005

10. Granger S. Hypnotherapy for childbirth. Pract Midwife. 2012;15(8):S13–4.

11. Oakley S, Evans E. The role of yoga: breathing, meditation and optimal fetal positioning. Pract Midwife. 2014;17(5):30–32.

12. Orme-Johnson D. Medical care utilization and the transcendental meditation program. Psychosom Med. 1987;49(5):493–507. doi:10.1097/00006842-198709000-00006

13. Domínguez-Solís E, Lima-Serrano M, Lima-Rodríguez JS. Non-pharmacological interventions to reduce anxiety in pregnancy, labour and postpartum: a systematic review. Midwifery. 2021;102:103126. doi:10.1016/j.midw.2021.103126

14. Thomson G, Feeley C, Moran VH, Downe S, Oladapo OT. Women’s experiences of pharmacological and non-pharmacological pain relief methods for labour and childbirth: a qualitative systematic review. Reprod Health. 2019;16(1):71. doi:10.1186/s12978-019-0735-4

15. Davis J. Effective non-pharmacological birth interventions. Pract Midwife. 2015;18(2):13–17.

16. Zeidan F, Vago DR. Mindfulness meditation–based pain relief: a mechanistic account. Ann NY Acad Sci. 2016;1373(1):114–127. doi:10.1111/nyas.13153

17. Zeidan F, Emerson NM, Farris SR, et al. Mindfulness meditation-based pain relief employs different neural mechanisms than placebo and sham mindfulness meditation-induced analgesia. J Neurosci. 2015;35(46):15307–15325. doi:10.1523/JNEUROSCI.2542-15.2015

18. Zeidan F, Adler-Neal AL, Wells RE, et al. Mindfulness-meditation-based pain relief is not mediated by endogenous opioids. J Neurosci. 2016;36(11):3391–3397. doi:10.1523/JNEUROSCI.4328-15.2016

19. Wang R, Lu J, Chow KM. Effectiveness of mind–body interventions in labour pain management during normal delivery: a systematic review and meta-analysis. Int J Nurs Studies. 2024;158:104858. doi:10.1016/j.ijnurstu.2024.104858

20. Smith CA, Levett KM, Collins CT, Dahlen HG, Ee CC, Suganuma M. Massage, reflexology and other manual methods for pain management in labour. Cochrane Database Syst Rev. 2018;2018(3). doi:10.1002/14651858.CD009290.pub3

21. Smith CA, Levett KM, Collins CT, Armour M, Dahlen HG, Suganuma M. Relaxation techniques for pain management in labour. Cochrane Database Syst Rev. 2018;2018(3). doi:10.1002/14651858.CD009514.pub2

22. Organization WH. WHO recommendations on intrapartum care for a positive childbirth experience: World Health Organization; 2018.

23. Orellana-Rios CL, Radbruch L, Kern M, et al. Mindfulness and compassion-oriented practices at work reduce distress and enhance self-care of palliative care teams: a mixed-method evaluation of an “on the job” program. BMC Palliative Care. 2017;17(1):3. doi:10.1186/s12904-017-0219-7

24. Mumtaz N, Tariq MZ, Ali A, et al. Impact of relaxation techniques on anxiety and pain management during labor: a quasi-experimental study. Pak-Euro J Med Life Sci. 2024;7(3):475–482.

25. Kabat-Zinn J. An outpatient program in behavioral medicine for chronic pain patients based on the practice of mindfulness meditation: theoretical considerations and preliminary results. General Hospital Psychiatry. 1982;4(1):33–47. doi:10.1016/0163-8343(82)90026-3

26. Duncan LG, Cohn MA, Chao MT, Cook JG, Riccobono J, Bardacke N. Benefits of preparing for childbirth with mindfulness training: a randomized controlled trial with active comparison. BMC Pregnancy Childbirth. 2017;17(1):140. doi:10.1186/s12884-017-1319-3

27. Başgöl Ş, Koç E. Non-pharmacological techniques in labor pain management. Samsun Saglik Bilimleri Dergisi. 2020;5(1):1–5.

28. Bartlett L, Martin A, Neil AL, et al. A systematic review and meta-analysis of workplace mindfulness training randomized controlled trials. J Occup Health Psychol. 2019;24(1):108. doi:10.1037/ocp0000146

29. Sandiford R. Keeping it natural. Nurs Times. 2006;102(3):22–23.

30. Young D. What is normal childbirth and do we need more statements about it? Birth. 2009;36(1):1–3. doi:10.1111/j.1523-536X.2008.00306.x

31. Ayres-de-Campos D, Spong CY, Chandraharan E. FIGO intrapartum fetal monitoring expert consensus panel. FIGO consensus guidelines on intrapartum fetal monitoring: cardiotocography. Int J Gynaecol Obstet. 2015;131(1):13–24. doi:10.1016/j.ijgo.2015.06.020

32. Zhang D, Tsang KW, Duncan LG, et al. Effects of the Mindfulness-Based Childbirth and Parenting (MBCP) program among pregnant women: a randomized controlled trial. Mindfulness. 2023;14(1):50–65. doi:10.1007/s12671-022-02046-8

33. Mansfield B. The social nature of natural childbirth. Soc Sci Med. 2008;66(5):1084–1094. doi:10.1016/j.socscimed.2007.11.025

34. Azizmohammadi S, Azizmohammadi S. Hypnotherapy in management of delivery pain: a review. Eur J Transl Myol. 2019;29(3):8365. doi:10.4081/ejtm.2019.8365

35. Evans MI, Britt DW, Worth J, et al. Uterine contraction frequency in the last hour of labor: how many contractions are too many? J Matern Fetal Neonatal Med. 2022;35(25):8698–8705. doi:10.1080/14767058.2021.1998893

36. Vargas-Uricoechea H, Castellanos-Pinedo A, Urrego-Noguera K, et al. Mindfulness-based interventions and the hypothalamic–pituitary–adrenal axis: a systematic review. Neurol Int. 2024;16(6):1552–1584. doi:10.3390/neurolint16060115

37. Ring HZ, Kern RJH. Zen meditation and the neuro-immuno-endocrine axis. Health. 2024;16(12):1242–1249. doi:10.4236/health.2024.1612086

38. Pascoe MC, Thompson DR, Ski CF. Metabolism: meditation and endocrine health and wellbeing. Trends Endocrinol Metab. 2020;31(7):469–477. doi:10.1016/j.tem.2020.01.012

39. Peternelj-Taylor C. Pregnancy, childbirth, and mothering: a forensic nursing response. J Forensic Nurs. 2008;4(2):53–54. doi:10.1111/j.1939-3938.2008.00009.x

40. Ayers S. Fear of childbirth, postnatal post-traumatic stress disorder and midwifery care. Midwifery. 2014;30(2):145–148. doi:10.1016/j.midw.2013.12.001

41. Li L, Zhang P, Qin Z, et al. The effect of holographic Meridian scraping therapy combined with free position on the labor process, perineum lateral resection rate, and delivery outcomes of primiparae. Am J Transl Res. 2021;13(8):9846–9852.

42. Guo L, Chen L, Jiao Y, et al. Analysis of the effect of free position delivery on the success rate and safety of vaginal trial delivery in patients with scar uterine vaginal delivery. Panminerva Med. 2022;64(4):574–576. doi:10.23736/S0031-0808.21.04478-5

43. Dénakpo J, Lokossou A, Tonato-Bagnan JA, et al. [Delivery in free position perhaps a solution to change delivery in traditional position in delivery rooms in Africa: results of a prospective study in Cotonou in Bénin]. J Obstet Gynaecol Can. 2012;34(10):947–953. doi:10.1016/S1701-2163(16)35408-1

44. Veringa-Skiba IK, Ziemer K, de Bruin EI, et al. Mindful awareness as a mechanism of change for natural childbirth in pregnant women with high fear of childbirth: a randomised controlled trial. BMC Pregnancy Childbirth. 2022;22(1):47. doi:10.1186/s12884-022-04380-0

45. Hughes A, Williams M, Bardacke N, et al. Mindfulness approaches to childbirth and parenting. Br J Midwifery. 2009;17(10):630–635. doi:10.12968/bjom.2009.17.10.44470

Introduction

Type 2 diabetes mellitus (T2DM) is a chronic metabolic disorder characterized by insulin resistance, hyperglycaemia, and various metabolic abnormalities.¹ Over the past few decades, the global prevalence of T2DM has continuously increased at an alarming rate, with estimates suggesting that more than 463 million people worldwide are currently living with diabetes, the vast majority of which are T2DM cases.² This surge is largely driven by rising obesity rates, unhealthy lifestyles, and population aging.³ T2DM is not only a major cause of morbidity and mortality but also places a tremendous economic burden on healthcare systems due to the frequent development of complications, including cardiovascular disease, nephropathy, and neuropathy.⁴

Among the numerous complications associated with T2DM, abnormal fat distribution plays a pivotal role in the progression of the disease and the development of insulin resistance.⁵ Visceral fat is known to contribute significantly to conditions related to metabolic dysfunction, such as dyslipidaemia, hypertension, and inflammation. Subcutaneous fat, though less harmful than visceral fat, also can contribute to metabolic derangements when present in excess.⁶ Given the importance of fat distribution in the pathophysiology of T2DM, understanding the factors that influence fat storage and distribution is critical for the development of effective therapeutic interventions.⁷

Thyroid hormones are well-established regulators of metabolism and energy expenditure.⁸ These hormones, including free triiodothyronine (FT3) and free thyroxine (FT4), play key roles in lipid metabolism, thermogenesis, and glucose homeostasis. Thyroid-stimulating hormone (TSH), which is produced by the pituitary gland, regulates the synthesis and secretion of both FT3 and FT4.^9,10 In clinical practice, thyroid function is typically evaluated by measuring thyroid hormone levels, with deviations from normal levels indicating either hypo- or hyperthyroidism.^11,12 Interestingly, research has demonstrated that even in euthyroid individuals (those with normal thyroid function), variations in thyroid hormone levels may have significant metabolic consequences.^13,14

Several studies^15–17 have explored the interplay between thyroid hormones and fat distribution, particularly in relation to obesity and metabolic syndrome. One study¹⁸ found that individuals with elevated FT3 levels tend to have higher body fat percentages, particularly involving visceral fat, which is associated with greater risks of insulin resistance and cardiovascular diseases. Conversely, higher FT4 levels are linked to lower body fat and improved metabolic profiles, suggesting a protective role of FT4 in fat metabolism. These findings are consistent with the known metabolic effects of thyroid hormones, with FT3 being the more metabolically active hormone that increases the basal metabolic rate and promotes lipid oxidation.¹⁹

However, the relationship between thyroid hormone levels and fat distribution in T2DM patients remains poorly understood. Most studies of this relationship have focused on general populations or individuals with obesity, without specifically examining T2DM patients. Given the unique metabolic characteristics of T2DM patients, including insulin resistance and altered glucose metabolism, it is important to investigate how thyroid hormone levels influence fat distribution in this population.

The present study aimed to elucidate the relationship between thyroid hormones (FT3, FT4, and TSH) and fat distribution, specifically in terms of visceral fat area (VFA) and subcutaneous fat area (SFA), in T2DM patients with normal thyroid function. By characterizing these relationships, a goal of this study was to provide insight into the potential roles of thyroid hormones in the pathophysiology of T2DM and its complications. This information, in turn, can support the development of targeted treatments for metabolic management in T2DM patients.

Materials and Methods

Study Design and Participants

This cross-sectional study included 2,956 T2DM patients recruited from a tertiary hospital between 2021 and 2023. The initial dataset contained 10,416 entries. The inclusion criteria were: (1) age ≥18 years, (2) confirmed diagnosis of T2DM according to the World Health Organization (WHO) criteria, and (3) availability of complete data for VFA, SFA, and thyroid hormone levels (FT3, FT4, and TSH). The exclusion criteria were applied as follows: duplicate entries (6,237 cases), missing age or age <18 years (7 cases), incomplete VFA or SFA data (875 cases), unconfirmed T2DM diagnosis (31 cases), presence of thyroid disease (122 cases), and missing FT3/FT4/TSH data (188 cases). After these exclusions, 2,956 patients with complete data were included in the final analysis.

The study was approved by the Institutional Review Board of the hospital, and written informed consent was obtained from all participants prior to enrolment.

Data Collection

Demographic and Clinical Characteristics

Baseline demographic and clinical data were collected from medical records and included: age, gender, body mass index (BMI), and waist-to-hip ratio (WHR); blood pressure (systolic and diastolic); and lifestyle factors such as smoking status (never, quit, occasionally, or daily) and alcohol consumption (never, quit, occasionally, or daily).

Thyroid Hormone Measurements

Fasting blood samples were obtained in the morning from all participants for the measurement of thyroid hormone levels.²⁰ Serum FT3, FT4, and TSH levels were measured using enzyme-linked immunosorbent assays (ELISAs). To explore differences in fat distribution across varying levels of thyroid hormones, we stratified participants into three groups (low, normal, and high), based on serum FT3, FT4, and TSH levels. The cutoffs for these categories were defined according to standard clinical reference ranges used in routine laboratory practice and prior studies on thyroid function in euthyroid individuals: FT3 (low <1.8, normal 1.8–3.8, high >3.8 pg/mL), FT4 (low <0.78, normal 0.78–1.86, high >1.86 ng/dL), and TSH (low <0.38, normal 0.38–5.57, high >5.57 mU/L). These ranges were chosen to ensure consistency with clinical definitions and to assess potential trends across the spectrum of thyroid hormone concentrations within the euthyroid range. Quality control procedures were applied to ensure measurement accuracy, and extreme values were addressed based on statistical feedback. Specifically, values for creatinine (Cr), FT4, and TSH were adjusted as follows: Cr (151.72 μmol/L and 103.9 μmol/L), FT4 (7.71 ng/dL), and TSH (97.743 mU/L). These adjustments corrected data entry inconsistencies or biologically implausible values, improving data integrity and minimizing potential bias. Sensitivity analyses indicated these corrections did not materially alter the study’s main findings.

Fat Distribution Assessment

Fat distribution was assessed by measuring VFA and SFA via bioelectrical impedance analysis (BIA)²¹ using the DUALSCAN HDS-2000 instrument (Omron Healthcare, Kyoto, Japan). Measurements were taken after an overnight fast and voiding of the bladder. The device calculates body impedance through weak electrical currents passing through the body, and participant data (hospital number, height, weight, age, gender) were entered for precise calculations. The abdominal measurement unit on the device measured abdominal shape, and impedance measurements allowed for calculation of VFA and SFA according to the manufacturer’s protocols.

Statistical Analysis

Descriptive Statistics

The baseline characteristics of the study population are presented as means ± standard deviations for normally distributed continuous variables, medians (interquartile ranges) for skewed variables, and as frequencies and percentages for categorical variables. Differences in baseline characteristics were assessed using the t-test for normally distributed continuous variables, the Mann–Whitney U-test for skewed variables, and the Chi-square test for categorical variables. Skewed variables (eg, triglycerides [TG], alanine transaminase [ALT]) were analyzed without log transformation. The level of statistical significance was set at P<0.05.

Correlation Analysis

Correlations between levels of thyroid hormones (FT3, FT4, TSH) and fat distribution measures (VFA, SFA) were analyzed via Pearson or Spearman correlation analysis based on the distribution of the data. Correlation results were evaluated at a significant level of P<0.05.

Multivariate Regression Analysis

Multiple linear regression models were employed to evaluate the independent effects of FT3, FT4, and TSH on VFA and SFA after adjusting for potential confounding variables, including age, gender, and BMI. Based on multicollinearity analyses, BMI and WHR were excluded from final models involving FT4 to avoid obscuring its significance in relation to fat distribution. Regression coefficients, confidence intervals, and P-values were reported for all predictors. All analyses were conducted using a statistical software package with significance set at P<0.05.

Handling of Outliers

Extreme values identified for Cr, FT4, and TSH measurements were corrected based on statistical review and clinical evaluation to enhance the robustness of the findings. A total of five values were adjusted: two for Cr (151.72 and 103.9 μmol/L), one for FT4 (7.71 ng/dL), and two for TSH (97.743 mU/L and one additional unreported outlier). These corrections addressed data entry inconsistencies or biologically implausible values, thus improving data quality and minimizing potential bias. Sensitivity analyses performed to assess the impact of these corrections showed no material changes in the direction, magnitude, or statistical significance of the associations between thyroid hormone levels and fat distribution. All statistical analyses in this study were performed using SPSS version 28.0 (IBM Corp., Armonk, NY, USA) and R software (version 4.2.2, R Foundation for Statistical Computing, Vienna, Austria), with results considered statistically significant at P<0.05.

Results

Baseline Characteristics of the Study Population

The baseline characteristics of the study participants (n=2,956) are summarized in Table 1, with stratification by gender. The mean age of the participants was 56.2 years, and the study population included a higher proportion of males (58.6%) than females (41.4%). Males had a significantly higher WHR and systolic blood pressure than females (both P<0.001). Lifestyle factors also showed gender-specific differences, with males reporting higher smoking and alcohol consumption rates than females. BMI did not differ significantly between the male and female groups (P=0.077).

Table 1 The Basic Characteristics Stratified by Gender

Thyroid Hormone Levels and Fat Distribution in T2DM Patients

Table 2 presents the distribution of VFA and SFA across groups of patients with different categories of thyroid hormone levels (decreased, normal, and increased) for FT3, FT4, and TSH. Notably, patients with elevated FT3 levels (>3.8 pg/mL) show higher mean values for VFA (79.75 ± 40.12 cm²) and SFA (148.96 ± 57.06 cm²) compared to those with normal (1.8–3.8 pg/mL) or decreased (<1.8 pg/mL) FT3 levels (all P<0.001). Conversely, no significant variation in VFA and SFA values was observed for patients in different FT4 categories (decreased, normal, increased), suggesting a lesser impact of FT4 alone on fat distribution (P>0.05). Upon analysis according to the TSH categories, slightly higher values of VFA (82.39 ± 49.75 cm²) and SFA (162.44 ± 61.38 cm²) were observed in the elevated TSH group (>5.57 mU/L) compared with the decreased and normal TSH groups (all P<0.05 for both VFA and SFA).

Table 2 The Basic Characteristics Stratified by Age

Correlation Between Thyroid Hormone Levels and Fat Distribution in T2DM Patients

Correlation analyses revealed significant relationships between thyroid hormone levels and fat distribution parameters (Table 3). FT3 was positively correlated with both VFA (r=0.248, P<0.001) and SFA (r=0.190, P<0.001), indicating that a higher FT3 level is associated with greater fat accumulation. FT4, on the other hand, tended to show a weak inverse correlation with both VFA (r=–0.018, P=0.321) and SFA (r=–0.019, P=0.283), although these relationships were not statistically significant. TSH showed a weak but significant positive correlation with both VFA (r=0.064, P=0.001) and SFA (r=0.072, P<0.001), indicating an association with fat distribution but a weaker association than that observed between FT3 and fat distribution.

Table 3 The Distribution of VFA and SFA According to Thyroid Hormones

Correlation Between Thyroid Hormone Levels and Fat Distribution

Correlation analyses revealed significant associations between thyroid hormone levels and fat distribution parameters (Table 4). FT3 was positively correlated with both VFA (r = 0.248, P < 0.001) and SFA (r = 0.190, P < 0.001), indicating that higher FT3 levels were associated with greater visceral and subcutaneous fat accumulation. FT4 showed weak negative correlations with VFA (r = –0.018, P = 0.321) and SFA (r = –0.019, P = 0.283), but these associations were not statistically significant. TSH demonstrated weak yet significant positive correlations with both VFA (r = 0.064, P = 0.001) and SFA (r = 0.072, P < 0.001), although the strength of these relationships was less pronounced compared with FT3.

Table 4 The Correlation Between Thyroid Hormones and VFA and SFA

Predictive Values of Thyroid Hormones for Abnormal Fat Distribution

To further evaluate the independent effects of thyroid hormones on fat distribution, multivariate linear regression models were constructed after adjusting for potential confounding factors, including age, gender, and BMI (Table 5). FT3 remained independently and positively associated with both VFA (Coef = 8.06, 95% CI: 5.87–10.24, P < 0.001) and SFA (Coef = 9.71, 95% CI: 6.62–12.79, P < 0.001). In contrast, FT4 was inversely associated with VFA (Coef = –12.20, 95% CI: –17.36 to –7.03, P < 0.001) and SFA (Coef = –16.68, 95% CI: –23.98 to –9.38, P < 0.001), suggesting a potential protective role of FT4 against abnormal fat accumulation. TSH did not show significant associations with either VFA or SFA (all P > 0.05), indicating a relatively minor impact compared with FT3 and FT4.

Table 5 The Multivariate Linear Regression Between Thyroid Hormones and VFA and SFA

Gender and Age-Specific Variations in Fat Distribution Among T2DM Patients

Additional analyses with patient stratification by age and gender were conducted to further elucidate the relationships between thyroid hormone levels and fat distribution in T2DM patients. As shown in Table S1, the associations between FT3 and fat distribution parameters were consistent across age groups, with younger and middle-aged adults showing similar patterns of correlation for FT3 with VFA and SFA. Among women, FT4 exhibited a more pronounced inverse relationship with VFA, and this effect was not observed in men. Gender differences were also evident in the correlation of TSH and fat distribution, with men showing a stronger association of TSH with VFA and SFA compared with women (P<0.05 for interaction).

Discussion

The present investigation of the associations between thyroid hormone levels (FT3, FT4, and TSH) and fat distribution, specifically VFA and SFA, in T2DM patients revealed the following significant associations: FT3 was positively correlated with both VFA and SFA, while FT4 was inversely associated with these fat distribution measures. TSH displayed a weak positive association that did not persist after adjustment for confounding variables. These results suggest a nuanced role for thyroid hormones in fat metabolism and distribution in T2DM patients with normal thyroid function, supporting a potential role for FT3 and FT4 as biomarkers for evaluating metabolic risk in this population.

The strong positive associations between FT3 and both VFA and SFA in our cohort align with FT3’s known role as an active thyroid hormone involved in regulating metabolic rate and energy expenditure.²² FT3’s influence on lipolysis and fat accumulation may explain the observed associations, as higher FT3 levels likely stimulate greater metabolic turnover, promoting fat mobilization and deposition, particularly in visceral regions.²³ The accumulation of visceral fat is closely linked to insulin resistance, systemic inflammation, and cardiovascular complications, all of which are prevalent in T2DM patients.²⁴ Our findings are consistent with those of prior studies^25,26 indicating that elevated FT3 levels are associated with higher body fat percentages and greater metabolic risks, especially in populations predisposed to obesity and diabetes.

The relationship between FT3 and fat distribution also could be influenced by the unique metabolic environment of T2DM.²⁷ Insulin resistance, a hallmark of T2DM, is associated with altered lipid metabolism and may exacerbate FT3’s effects on fat accumulation.²⁸ While FT3 promotes lipid oxidation and energy consumption, its effects in the context of T2DM may be compounded by impaired glucose and lipid regulation, resulting in disproportionate fat deposition, particularly in visceral areas.²⁹ Thus, T2DM patients with elevated FT3 levels may be at increased risk of accumulating metabolically active visceral fat, predisposing them to worsened metabolic control and cardiovascular risk. This highlights the potential utility of FT3 as a biomarker for assessing fat-related metabolic risks in T2DM.³⁰

The observed positive association between FT3 levels and fat accumulation may appear paradoxical considering the known lipolytic effects of FT3. However, previous research has suggested several mechanisms underlying this relationship in insulin-resistant states. Elevated FT3 levels in T2DM patients could reflect compensatory hypersecretion intended to overcome reduced thyroid hormone sensitivity at peripheral tissues,³⁰ potentially due to alterations in thyroid hormone receptor expression or signaling pathways. Additionally, impaired peripheral conversion of T4 to active T3 caused by chronic inflammation or metabolic dysfunction may lead to higher FT3 secretion from the thyroid gland to maintain metabolic homeostasis. These compensatory adjustments may paradoxically leads to the association of elevated circulating FT3 levels with greater fat accumulation, especially visceral fat, in individuals with insulin resistance and metabolic disturbances.³¹

Unlike FT3, FT4 displayed an inverse relationship with VFA and SFA, suggesting a potential protective role of FT4 in fat metabolism among T2DM patients.³² FT4, often considered less metabolically active than FT3, appears to exert a regulatory influence on fat storage, potentially attenuating fat accumulation in both visceral and subcutaneous regions.³³ This finding aligns with the hypothesis that FT4 supports metabolic homeostasis and thereby contributes to favorable fat distribution patterns in T2DM patients. Previous studies^14,34 have shown that lower FT4 levels are linked to increased fat mass and metabolic risk, further corroborating our finding that higher FT4 levels might protect against excessive fat accumulation.

Our gender-stratified analysis suggested that the protective effect of FT4 may be more pronounced in females than in males. A potential mechanism may be related to estrogen. Estrogen has been shown to influence fat distribution by favoring subcutaneous over visceral fat storage, and this effect may interact with the influence of FT4 on metabolism.³⁵ In postmenopausal women, who experience shifts in both thyroid hormone levels and fat distribution, maintaining adequate FT4 levels might help mitigate visceral fat accumulation.³⁶ This interplay warrants further investigation to explore whether FT4 could be targeted in gender-specific interventions aimed at improving fat distribution and metabolic outcomes in T2DM patients.

TSH demonstrated a weak positive correlation with VFA and SFA, which did not remain significant after adjustment for age, gender, and BMI.³⁷ As a regulatory hormone, TSH primarily influences thyroid hormone synthesis and secretion rather than directly impacting metabolism or fat storage. This may explain its limited association with fat distribution in our cohort. Although previous research³⁸ suggested a link between higher TSH levels and obesity, our results indicate that TSH may have minimal impact on fat distribution in euthyroid individuals with normal FT3 and FT4 levels. The lack of a significant association between TSH and fat accumulation after adjustment highlights the importance of considering FT3 and FT4 independently when assessing thyroid hormone-related metabolic risks in T2DM patients.

The relatively weak and statistically nonsignificant adjusted associations observed between TSH and fat distribution parameters may be explained by the role of TSH primarily as a centrally controlled regulator of thyroid function, rather than a direct peripheral mediator of metabolic processes. TSH exerts indirect metabolic effects by influencing thyroid hormone synthesis and secretion, whereas peripheral tissues primarily respond to circulating FT3 and FT4 levels. Furthermore, existing research supports the notion that peripheral metabolic activities and fat metabolism are predominantly regulated by thyroid hormones at the tissue level, with TSH having limited direct effects outside the central axis.³⁹ These findings align with our observation that FT3 and FT4 showed more robust associations with fat distribution compared with TSH.

Overall, our results are consistent with previous studies^40,41 that described a positive association of FT3 with body fat and an inverse relationship of FT4 with fat accumulation. Similar findings in general populations suggest that these thyroid hormone relationships extend to T2DM patients, who exhibit distinct metabolic profiles due to insulin resistance and dyslipidemia. Studies in non-diabetic populations have linked higher FT3 levels with increased visceral and subcutaneous fat, reinforcing our observations that FT3 may predispose T2DM patients to adverse fat accumulation patterns.⁴² Moreover, a study by Mele et al³⁶ demonstrated that FT4 levels are inversely correlated with adiposity, suggesting that FT4 might contribute to a healthier fat distribution pattern. However, our study’s focus on euthyroid T2DM patients provides novel insights and fills a gap in the literature, as few studies have exclusively examined this population.

As noted above, the significant associations between FT3 and measures of fat distribution suggest that FT3 could serve as a useful biomarker for identifying T2DM patients at risk for excessive visceral fat accumulation. Accordingly, monitoring of FT3 levels in T2DM patients may help clinicians identify individuals with heightened metabolic and cardiovascular risks linked to visceral adiposity. Moreover, our finding regarding the protective role of FT4 against abnormal fat distribution highlight the potential benefits of preserving or optimizing FT4 levels to promote a favorable fat distribution in T2DM patients. Personalized treatment approaches that consider individual thyroid hormone profiles could enhance metabolic management, potentially facilitating measures to reduce fat accumulation and improve metabolic outcomes in T2DM patients. Interventions focusing on lifestyle modification or pharmacotherapy tailored to regulate patients’ FT3 and FT4 levels could provide targeted benefits.

The observed gender-specific differences in the effect of FT4 on fat distribution underscore the importance of considering gender in metabolic research. Among the women in our study, we observed a more pronounced inverse relationship between FT4 and VFA, suggesting that FT4 may offer stronger protective effects in women. Estrogen likely interacts with FT4, influencing fat distribution patterns that favour subcutaneous over visceral storage. This interaction is particularly relevant for postmenopausal women, who experience hormonal shifts that might exacerbate visceral fat accumulation. Thus, strategies to maintain adequate FT4 levels could mitigate these effects, supporting metabolic health in postmenopausal T2DM patients. Future studies should explore the existence and effects of interactions between estrogen and thyroid hormones to better understand their combined effects on fat distribution in T2DM patients.

The present study also has several limitations that should be considered. The cross-sectional design prevents us from establishing causality, leaving it unclear whether thyroid hormone changes precede or result from altered fat distribution. A longitudinal study design is required to clarify causal pathways. Additionally, the exclusion of T2DM patients with thyroid dysfunction restricts the generalizability of our findings, as the relationship between thyroid hormones and fat distribution may differ in patients with hypo- or hyperthyroidism. The used of BIA for measurement of VFA and SFA may also limit the precision of the data, compared with that produced by imaging modalities such as MRI or CT, but BIA offers the advantage of being a practical, non-invasive method for large-scale studies. Residual confounding is another limitation. Although we controlled age, gender, BMI, education, blood pressure, and lipid profiles, other clinically relevant factors were not available in our dataset. Examples include the duration of diabetes, glycemic control (HbA1c), the use of specific antidiabetic medications, smoking status, and physical activity. These factors may influence both thyroid hormone metabolism and fat distribution, and their absence from our analysis may introduce bias. Therefore, residual confounding cannot be excluded and should be considered when interpreting our findings. Future studies should aim to account for these variables to provide a more comprehensive understanding of the impacts of different thyroid hormones on fat distribution.

Finally, this study provides insight for the directions of future research. Longitudinal investigations tracking thyroid hormone levels and fat distribution over time are needed to clarify potential causal relationships in T2DM populations. Complementary mechanistic studies, including observational analyses and experimental models, may help determine whether variations in FT3 and FT4 drive fat accumulation or represent compensatory responses. Elucidating the molecular pathways linking thyroid hormones with lipid metabolism will be particularly important. In addition, examination of other thyroid hormone metabolites, such as T2 and reverse triiodothyronine (rT3), may broaden our understanding of thyroid hormone regulation of fat distribution and metabolic health.

Conclusion

In conclusion, this study highlights significant associations between thyroid hormones and fat distribution in euthyroid T2DM patients. Elevated FT3 levels were linked to increased visceral and subcutaneous fat, whereas higher FT4 levels were associated with more favourable fat distribution. These findings provide new insights into the metabolic roles of thyroid hormones and suggest that they may serve as useful indicators for identifying patients at higher metabolic risk. Future longitudinal and mechanistic studies are warranted to validate these associations and to explore their potential relevance for personalized metabolic management in T2DM.

Data Sharing Statement

The datasets generated and analyzed during the present study are available from the corresponding author on reasonable request.

Ethics Approval and Informed Consent

The study was approved by the Institutional Review Board of the Affiliated Hospital of Southwest Medical University. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Written informed consent was obtained from all participants prior to enrolment.

Acknowledgments

We thank Medjaden, Inc. for scientific editing of this paper.

Funding

The study was funded by the Youth Fund of Medical Association (No. 16213).

Disclosure

The authors report no conflicts of interest in this work.

References

1. Hurtado MD, Vella A. What is type 2 diabetes? Medicine. 2019;47(1):10–15. doi:10.1016/j.mpmed.2018.10.010

2. Abdul Basith Khan M, Hashim MJ, King JK, Govender RD, Mustafa H, Al Kaabi J. Epidemiology of type 2 diabetes — global burden of disease and forecasted trends. J Epidemiol Global Health. 2020;10(1):107–111. doi:10.2991/jegh.k.191028.001

3. Saeedi P, Petersohn I, Salpea P, et al. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: results from the International Diabetes Federation Diabetes Atlas, 9th edition. Diabetes Res Clin Pract. 2019;157:107843. doi:10.1016/j.diabres.2019.107843

4. Zakir M, Ahuja N, Surksha MA, et al. Cardiovascular complications of diabetes: from microvascular to macrovascular pathways. Cureus. 2023;15(9):e45835. doi:10.7759/cureus.45835

5. Chandrasekaran P, Weiskirchen R. The Role of Obesity in Type 2 Diabetes Mellitus—An Overview. Int J Mol Sci. 2024;25(3):1882. doi:10.3390/ijms25031882

6. Goswami B, Reang T, Sarkar S, Sengupta S, Bhattacharjee B. Role of body visceral fat in hypertension and dyslipidemia among the diabetic and nondiabetic ethnic population of Tripura—A comparative study. J Family Med Primary Care. 2020;9(6).

7. Ruze R, Liu T, Zou X, et al. Obesity and type 2 diabetes mellitus: connections in epidemiology, pathogenesis, and treatments. Front Endocrinol. 2023;14.

8. Yavuz S, Salgado Nunez Del Prado S, Celi FS. Thyroid hormone action and energy expenditure. J Endocrine Soc. 2019;3(7):1345–1356. doi:10.1210/js.2018-00423

9. Wang X, Wu Z, Liu Y, et al. The role of thyroid-stimulating hormone in regulating lipid metabolism: implications for body–brain communication. Neurobiol Dis. 2024;201:106658. doi:10.1016/j.nbd.2024.106658

10. Bansal S, Vachher M, Arora T, Kumar B, Burman A. Visceral fat: a key mediator of NAFLD development and progression. Hum Nutr Metab. 2023;33:200210. doi:10.1016/j.hnm.2023.200210

11. Sterenborg RBTM, Steinbrenner I, Li Y, et al. Multi-trait analysis characterizes the genetics of thyroid function and identifies causal associations with clinical implications. Nat Commun. 2024;15(1):888. doi:10.1038/s41467-024-44701-9

12. Jonklaas J, Bianco AC, Bauer AJ, et al. Guidelines for the treatment of hypothyroidism: prepared by the American Thyroid Association Task Force on thyroid hormone replacement. Thyroid. 2014;24(12):1670–1751. doi:10.1089/thy.2014.0028

13. Cappola AR, Arnold AM, Wulczyn K, Carlson M, Robbins J, Psaty BM. Thyroid function in the euthyroid range and adverse outcomes in older adults. J Clin Endocrinol Metab. 2015;100(3):1088–1096. doi:10.1210/jc.2014-3586

14. Gokkaya N, Aydin K. Efficacy of levothyroxine monotherapy in achieving clinical euthyroidism and its impact on weight loss in women with hypothyroidism and obesity. Scientific Reports. 2024;14(1):27822. doi:10.1038/s41598-024-78185-w

15. Kim JM, Kim BH, Lee H, et al. The relationship between thyroid function and different obesity phenotypes in Korean euthyroid adults. Diabetes Metab J. 2019;43(6):867–878. doi:10.4093/dmj.2018.0130

16. Teixeira PFS, Dos Santos PB, Pazos-Moura CC. The role of thyroid hormone in metabolism and metabolic syndrome. Therapeu Adv Endocrinol Metabo. 2020;11:2042018820917869. doi:10.1177/2042018820917869

17. Abiri B, Ahmadi AR, Mahdavi M, Amouzegar A, Valizadeh M. Association between thyroid function and obesity phenotypes in healthy euthyroid individuals: an investigation based on Tehran Thyroid Study. Eur J Med Res. 2023;28(1):179. doi:10.1186/s40001-023-01135-1

18. Adamska A, Raczkowski A, Stachurska Z, et al. Body composition and serum concentration of thyroid hormones in euthyroid men and women from general population. J Clin Med. 2022;11(8):2118. doi:10.3390/jcm11082118

19. Hatziagelaki E, Paschou SA, Schön M, Psaltopoulou T, Roden M. NAFLD and thyroid function: pathophysiological and therapeutic considerations. Trends Endocrinol Metab. 2022;33(11):755–768. doi:10.1016/j.tem.2022.08.001

20. Ward CR. Chapter 69 – Thyroid Storm. In: Silverstein DC, Hopper K, editors. Small Animal Critical Care Medicine. 2nd ed. St. Louis: W.B. Saunders; 2015:364–367.

21. Lai C-L, Lu H-K, Huang A-C, Chu L-P, Chuang H-Y, Hsieh K-C. Bioimpedance analysis combined with sagittal abdominal diameter for abdominal subcutaneous fat measurement. Front Nutr. 2022;9.

22. Hoermann R, Pekker MJ, Midgley JEM, Dietrich JW. The role of supporting and disruptive mechanisms of FT3 homeostasis in regulating the hypothalamic–pituitary–thyroid axis. Therapeu Adv Endocrinol Metabo. 2023;14:20420188231158163. doi:10.1177/20420188231158163

23. Hu Y, Zhou F, Lei F, et al. The nonlinear relationship between thyroid function parameters and metabolic dysfunction-associated fatty liver disease. Front Endocrinol. 2023;14.

24. Rohm TV, Meier DT, Olefsky JM, Donath MY. Inflammation in obesity, diabetes, and related disorders. Immunity. 2022;55(1):31–55. doi:10.1016/j.immuni.2021.12.013

25. Sun X, Chen L, Wu R, Zhang D, He Y. Association of thyroid hormone with body fat content and lipid metabolism in euthyroid male patients with type 2 diabetes mellitus: a cross-sectional study. BMC Endocr Disord. 2021;21(1):241. doi:10.1186/s12902-021-00903-6

26. Yu L, Liu Y, Wang Y, et al. Increased thyroid hormone sensitivity is correlated with visceral obesity in patients with type 2 diabetes. Lipids health Dis. 2024;23(1):337. doi:10.1186/s12944-024-02320-9

27. Bouazza A, Favier R, Fontaine E, Leverve X, Koceir E-A. Potential applications of thyroid hormone derivatives in obesity and type 2 diabetes: focus on 3,5-diiodothyronine (3,5-T2) in psammomys obesus (fat sand rat) model. Nutrients. 2022;14(15):3044. doi:10.3390/nu14153044

28. DeFronzo RA. Insulin resistance, lipotoxicity, type 2 diabetes and atherosclerosis: the missing links. The Claude Bernard Lecture 2009. Diabetologia. 2010;53(7):1270–1287. doi:10.1007/s00125-010-1684-1

29. Emiroğlu C, Özdemir Ç, Görpelioğlu S, Aypak C. The relationship between thyroid hormones, metabolic syndrome and HOMA-IR in people with obesity or overweight. Clin Diabetol. 2022;11(5):333–339. doi:10.5603/DK.a2022.0043

30. Biondi B, Kahaly GJ, Robertson RP. Thyroid dysfunction and diabetes mellitus: two closely associated disorders. Endocrin Rev. 2019;40(3):789–824. doi:10.1210/er.2018-00163

31. Holtorf K. Peripheral thyroid hormone conversion and its impact on TSH and metabolic activity. J Restorat Med. 2014;3(1):30–52. doi:10.14200/jrm.2014.3.0103

32. Cao B, Li K, Ke J, Zhao D. Impaired sensitivity to thyroid hormones is associated with the change of abdominal fat in euthyroid type 2 diabetes patients: a retrospective cohort study. J Diabe Res. 2024;2024(1):8462987. doi:10.1155/2024/8462987

33. Freedland ES. Role of a critical visceral adipose tissue threshold (CVATT) in metabolic syndrome: implications for controlling dietary carbohydrates: a review. Nutr Metab. 2004;1(1):12.

34. Sanyal D, Raychaudhuri M. Hypothyroidism and obesity: an intriguing link. Indian J Endocrinol Metab. 2016;20(4):554–557. doi:10.4103/2230-8210.183454

35. Mazza E, Troiano E, Ferro Y, et al. Obesity, dietary patterns, and hormonal balance modulation: gender-specific impacts. Nutrients. 2024;16(11):1629. doi:10.3390/nu16111629

36. Mele C, Mai S, Cena T, et al. The pattern of TSH and fT4 levels across different BMI ranges in a large cohort of euthyroid patients with obesity. Front Endocrinol. 2022;13.

37. Shaoba A, Basu S, Mantis S, Minutti C. Serum thyroid-stimulating hormone levels and body mass index percentiles in children with primary hypothyroidism on levothyroxine replacement. J Clin Res Pediatr Endocrinol. 2017;9(4):337–343. doi:10.4274/jcrpe.3661

38. Witte T, Völzke H, Lerch MM, et al. Association between serum thyroid-stimulating hormone levels and visceral adipose tissue: a population-based study in Northeast Germany. Euro Thyroid J. 2017;6(1):12–19. doi:10.1159/000450977

39. McAninch EA, Bianco AC. Thyroid hormone signaling in energy homeostasis and energy metabolism. Ann N Y Acad Sci. 2014;1311:77–87. doi:10.1111/nyas.12374

40. Biondi B. Thyroid and obesity: an intriguing relationship. J Clin Endocrinol Metab. 2010;95(8):3614–3617. doi:10.1210/jc.2010-1245

41. Mullur R, Liu -Y-Y, Brent GA. Thyroid hormone regulation of metabolism. Physiol Rev. 2014;94(2):355–382. doi:10.1152/physrev.00030.2013

42. Tacke F, Horn P, Wai-Sun Wong V, et al. EASL–EASD–EASO clinical practice guidelines on the management of metabolic dysfunction-associated steatotic liver disease (MASLD). J Hepatol. 2024;81(3):492–542.

Following 2 successful workshops on AI in the energy sector, held in September 2025, the European Commission organises a third workshop on the use of AI in the renewables and buildings sectors. 

While AI holds immense potential to enhance manufacturing, optimisation, and improving maintenance, several challenges hinder effective implementation. These include data silos, lack of collaboration, and varying business models. The workshop will explore the role of the Commission in facilitating the swift and safe deployment of AI technologies in these sectors, discuss the ongoing work in the matter and inform about policy actions of the future Strategic Roadmap for Digitalisation and AI in the energy sector. 

The event is addressing stakeholders from digital and energy value chains, including technology providers, IT suppliers, SMEs, aggregators, system integrators, digital solutions providers, data centre operators, cloud service providers, consumers, energy communities, appliance manufacturers, research community, energy intensive industries, building operators, car manufacturers and providers of e-mobility solutions. 

The draft agenda will follow shortly but it is already possible to register.

Related links