Blog

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