Blog

Introduction

The Global Initiative for Chronic Obstructive Lung Disease (GOLD) releases an annual report outlining recommendations for diagnosis, management, and prevention of Chronic Obstructive Pulmonary Disease (COPD).¹ While the GOLD reports are formally classified as a strategy, they are commonly regarded as guidelines for prescription of inhaled medication. In the Netherlands, COPD care is guided by either the GOLD strategy or the national guideline, which largely reflect the GOLD recommendations regarding pharmacological treatment.^2,3 The 2023 edition of the GOLD strategy introduced significant updates to inhaled medication treatment.^4,5 These revised guidelines should have substantial implications for clinical practice. Some treatment recommendations remained consistent between the 2022 and 2023 strategy. Two treatment algorithms are used to select the appropriate follow-up inhaled therapy: the exacerbation algorithm for patients experiencing exacerbations and the dyspnoea algorithm for those with persistent dyspnoea. Furthermore, asthma as a comorbidity or history requires primary treatment for asthma, including the use of inhaled corticosteroids (ICS).^5,6 The 2023 strategy introduced three key changes compared with the 2022 recommendations: 1) the use of ICS in the absence of exacerbations, regardless of dyspnoea severity, is no longer recommended; 2) earlier escalation to triple therapy is advised for patients with elevated blood eosinophil counts (≥0.3 x10⁹/L), and 3) the combination of ICS and a long-acting beta-agonist (LABA) as a double therapy is no longer recommended. When ICS is indicated, triple therapy combining a LABA, a long-acting muscarinic antagonist (LAMA), and ICS is advised.⁵ This last recommendation was supported by two large trials that showed that triple therapy was more effective compared to LABA+ICS.^7,8 However, this change resulted in debate regarding patients who were clinically stable on LABA+ICS. To address this, the 2025 strategy introduced an important nuance: clinically stable patients already using LABA+ICS may continue this regimen.¹

There is considerable debate on the appropriate use of ICS in COPD. Although the GOLD strategy provides clear guidance on initiating ICS, it offers no recommendations on how or whether ICS should be withdrawn in (clinically stable) patients, aside from a brief note on discontinuation in case of side effects.^1,5,6 Overuse of ICS could potentially be harmful due to the increased risk for side effects (for example pneumonia). Conversely, there is an undesirable risk for undertreatment when withholding ICS in COPD patients with exacerbations and high eosinophil counts.⁹ Therefore, it is important to assess prescribing patterns in daily practice and to investigate clinicians’ adherence to the latest evidence-based recommendations. Clinicians may encounter barriers when implementing guideline recommendations in daily practice, for example due to resistance to modifying existing treatment regimens or frequent annual guideline updates.¹⁰ Identifying patients whose treatment more frequently deviates from the GOLD strategy may encourage clinicians to evaluate whether current treatment remains appropriate in these patients. Findings from this Dutch cohort study could provide valuable guidance for other healthcare systems facing similar challenges in implementing the GOLD recommendations.

Therefore, this descriptive study aims to evaluate clinicians’ adherence to the recommendations on inhaled medications of the GOLD strategy in daily practice and describe whether, and to what extent, the 2023 update has influenced prescribing patterns. Also, we aimed to identify subgroups of patients at risk for receiving inhaled medication not in line with GOLD recommendations. Furthermore, we explored ICS use and indication, specifically in the context of clinically stable patients. Lastly, barriers to implementation of the GOLD recommendations are discussed.

Methods

Study Design

In this retrospective observational cohort study, data of patients with COPD were collected from the Franciscus Gasthuis and Vlietland hospital, a large teaching hospital specializing in COPD and asthma care in Rotterdam, the Netherlands. First, data of patients with a COPD diagnosis in 2022 were automatically extracted from electronic patient records. Data on asthma as a comorbidity and lung function data that could not be automatically extracted were supplemented by the researchers manually. Secondly, to determine adherence to the guidelines, a cross-year comparison between 2022 and 2023 was conducted. Each patient’s treatment was assessed against the treatment algorithms in the 2022 GOLD strategy at the time it was last in effect. This assessment was repeated in November 2023, after the 2023 GOLD strategy had been in effect for one year. For adequate interpretation, it is important to note that the update of GOLD for the new year (eg “202x”) is routinely presented in the last months of the year before (202x-1). We assessed whether specific patient groups were less likely to receive treatment in accordance with the most recent recommendations. Finally, an in-depth evaluation of ICS use was performed. Ethics approval for this retrospective data study was waived by the Institutional Review Board of the Franciscus Gasthuis & Vlietland, Rotterdam, the Netherlands (identification number 2023–037). The study complies with the Declaration of Helsinki. This study was reported in line with the STROBE reporting checklist for observational studies.¹¹

Setting

Data were sourced from a secondary care hospital specializing in COPD and asthma. This hospital serves as a regional centre of expertise due to its coordination of care. Patients were referred to this hospital by their general practitioners or from other medical facilities, implying that initial therapy had likely been initiated elsewhere. Approximately 20 to 25 healthcare professionals, including treating physicians and specialized nurses, are responsible for the outpatient clinic care and therefore prescribe the inhaled medication.

Patient Selection

Data of all patients with an active COPD diagnosis and at least one inhaled medication prescription in the year before the release of the GOLD 2023 strategy (1–12-2021 until 01–12-2022) were included in the study. Patients were excluded from the study if essential data for objective clinical assessment were missing, such as eosinophil count or Clinical COPD Questionnaire (CCQ) score. Patients younger than 40 years were excluded to avoid including those with primarily asthma. Lastly, patients with a Forced Expiratory Volume in 1 second/Forced Vital Capacity (FEV1/FVC) ratio >0.7 were also excluded, as this criterion does not align with the GOLD definition of COPD.

Data Collection

Data on age, sex, inhaled medication, eosinophils, Clinical COPD Questionnaire (CCQ) score, exacerbations, hospitalizations, and the use of prednis(ol)one, azithromycin, or roflumilast, were automatically extracted from electronic patient records. An exacerbation was defined as a prescription of 30 or 40 mg prednis(ol)one daily for a duration of 3 to 14 days, which, according to COPD guidelines, is the primary treatment for COPD exacerbations in both primary and secondary care in the Netherlands. Use of other systemic corticosteroids is uncommon. The range of doses and number of days was chosen to capture potential variations in prescribing practices between physicians. Prednisone prescriptions were extracted from pharmacy records of our hospital and of the “Landelijk Schakel Punt (LSP)”. LSP is the national pharmacy data-sharing system in the Netherlands and includes the community prescriptions from general practitioners.¹² Antibiotics were excluded from the exacerbation definition as it is very uncommon in the Netherlands to treat COPD exacerbations with antibiotic monotherapy. Hospitalization was classified as any hospital stay linked to the COPD diagnosis, where the primary physician overseeing care was a pulmonologist. Asthma was considered a comorbidity next to COPD if the treating physician documented asthma, asthmatic characteristics, or ACO/ACOS in the medical history or conclusion in the patient record. However, if asthma was not mentioned or if only reversibility or bronchial hyperactivity were described, we considered this variable as negative. Most recent FEV1, FVC, and FEV1/FVC values before the visit to the pulmonologist were collected from the electronic patient record. If there was both a pre- and post-bronchodilator value known, the post value was used. Multiple eosinophil counts were extracted, including the most recent value, the highest count in the past year, the highest count over the past three years, and the highest count recorded ever. The CCQ is the widely used COPD questionnaire in the Netherlands, in contrast to COPD Assessment Test (CAT) or the Medical Research Council (MRC) dyspnoea scale. As the CCQ represents the same domains, these data were collected.¹³

Assessment of Adherence to Inhaled Medication Recommendations

We evaluated the inhaled medication at two time points: 1) the date of the most recent inhaled medication prescription before December 1, 2022, when the 2022 GOLD strategy applied; and 2) one year after the release of the 2023 strategy (November 14, 2023), Figure 1A. By that time, it could be assumed that patients had attended at least one follow-up visit since the new recommendations applied and, if necessary, had had the opportunity to have their inhaled medication adjusted. Our main outcome was the percentage of COPD patients with inhaled treatment according to the relevant GOLD strategy at each timepoint. The assessment of guideline adherence was based on the treatment algorithms in the GOLD strategies. Six objective parameters were used: (i) presence of asthma as comorbidity or in the past; (ii) number of exacerbations; (iii) number of hospitalizations; (iv) previous medication; (v) eosinophil count; and (vi) CCQ symptom score, Figure 1B-C. Guideline adherence was assessed by comparing the recommended treatment, as determined from the assessment shown in Figure 1, with the patient’s actual treatment. If these were identical, guideline adherence was scored as “adherent”; if not, it was scored as “nonadherent”. For the primary assessment of the treatment, we used the most recent CCQ symptom score, highest-ever eosinophil count, and the number of exacerbations and hospitalizations in the previous three years. The CCQ symptom score (the subdomain determined by the sum of question 1, 2, 5 and 6 dived by 4) was used as an objective parameter for dyspnoea. A threshold of 1.9 points was used for severe symptoms, based on a previous study.¹⁴ Our assessment followed the pharmacological follow-up protocols outlined in the GOLD strategies, rather than the initial treatment protocol.^5,6 Finally, while the follow-up inhaled medication protocol in the 2023 and currently applicable 2025 strategies is similar, the latter provides an additional nuance for stable patients on LABA/ICS. The 2025 strategy states that LABA/ICS could be continued if patients currently have no exacerbations, a low symptom load, and no features of asthma.¹ Therefore, we investigated how many patients in 2023 would have had treatment according to the guideline if the 2025 nuance was already applicable.

Figure 1 Overview of the 2022 and 2023 assessment. This figure provides an overview of the two assessments conducted in this study. (A) shows the timeline of the two assessments. (B) illustrates our assessment based on the follow-up treatment protocols in the 2022 GOLD strategy, while (C) illustrated the assessment based on the follow-up treatment protocols in the 2023 GOLD strategy. Abbreviations: LAMA, Long-Acting Muscarinic Antagonist; LABA, Long-Acting Beta-Agonist; ICS, Inhaled corticosteroids.

Differences Between Specific Patient Groups

To evaluate specific patient groups at higher risk of inadequate treatment, we used the 2023 strategy exclusively, since its recommendations are identical to those of the 2025 strategy currently in force. Guideline adherence was assessed in various subgroups: patients with low (<0.1 x10⁹/L), medium (0.1–0.3 x10⁹/L) and high (>0.3 x10⁹/L) eosinophil count; patients treated by the dyspnoea algorithm and those treated by the exacerbation algorithm; different GOLD classifications based on severity of airflow obstruction (I, II, III, IV); and patients on single inhaler triple therapy (SITT) or multiple inhaler triple therapy (MITT).

In-Depth Evaluation of ICS Use

To evaluate ICS use and its indications in greater depth, we first conducted a cross-sectional analysis at the second measurement point of this study (end of 2023), as the 2023 strategy recommendations for ICS are consistent with those of the 2025 strategy currently in effect. We then assessed whether adherence to the 2023 GOLD strategy changed when different timeframes for exacerbations and hospitalizations were applied (previous one, two, or three years). Additionally, we examined the impact of using either the most recent eosinophil count or the highest count recorded in the last year, last three years, or ever. Next, we compared the 2022 and 2023 strategy at the same historical time point: the end of 2022, which was the final moment the 2022 strategy was applicable. We determined ICS indications among ICS users according to the 2022 strategy and assessed what the indication would have been at that same time according to the 2023 strategy. Patients who had an indication under the 2022 strategy but not under the 2023 strategy represent those whose medication would need to be adjusted to align with the newest recommendations. These patients were then re-evaluated one year later, at the second measurement point of this study.

Statistical Analysis

Most outcomes are presented descriptively using means and standard deviations, median and interquartile ranges, or numbers and percentages, as appropriate. The primary outcome (guideline adherence at each time point) is presented as a percentage with the 95% confidence interval (CI). The McNemar test is used to test for statistical significant differences in guideline adherence between 2022 and 2023. The chi-square test was used to compare clinicians’ adherence to the GOLD strategy across different subgroups.

Results

Patient Characteristics

Data of 2434 patients with an active COPD diagnosis in 2022 were automatically extracted from the electronic patient records. After excluding patients under 40 years of age (n=5), those with FEV1/FVC >0.7 (n=85), and those with missing data on eosinophils (n=350) or CCQ score (n=675), a total of 1318 patients were included in the study cohort, Figure 2. Within this cohort, 689 (52.3%) patients were female, and the median age was 70 [interquartile range (IQR) 64–76], Table 1. Median FEV1 post bronchodilator percent predicted was 47 [IQR 36–63]. Asthma was documented as a comorbidity or past diagnosis in 245 patients (18.6%). The majority of patients in the study cohort (n=962, 73.0%) received triple therapy (the combination of a LABA, LAMA and ICS). This proportion of triple therapy users was higher in the study cohort than in the exclusion cohort based on missing eosinophils (54.0%) and the exclusion cohort based on missing CCQ (48.1%), Table S1. Consequently, LABA/LAMA and LABA/ICS use was lower in the study cohort (12.6% and 9.4% respectively) compared to the exclusion sub cohorts with missing eosinophils (21.1% and 11.4% respectively) and the sub cohort with missing CCQ (23.7% and 18.1% respectively). Maintenance use of prednisone and azithromycin was higher in the study cohort (4.7% and 10.9% respectively) than in the exclusion cohorts (2.0–2.1% and 3.0–3.7% respectively).

Figure 2 Patient selection.

Table 1 Baseline Characteristics

Assessment of Adherence to Inhaled Medication Recommendations

During the first measurement point of this study (end of 2022), 680 patients received inhaled medication in line with the recommendations of the 2022 strategy (51.6% with 95% CI interval 48.9–54.3%). During the second measurement point of this study (end 2023), one year after the 2022 analysis, 1074 patients of the study cohort were still alive (81.4%). Of those, the inhaled medication regimen of 615 patients was adherent to the 2023 strategy (57.3% with 95% CI interval 54.3–60.2%). There was no difference in guideline adherence to the 2022 strategy between patients who were alive in 2023 and those who had died: 51.8% and 50.8%, respectively (p=0.789).

Among the patients alive at both time points (n=1074), the inhaled medication prescriptions of 454 patients (42.3%) adhered to both GOLD strategies, 102 (9.5%) adhered in 2022 but not in 2023, 161 (15.0%) did not adhere in 2022 but did in 2023, and the prescriptions of 357 patients (33.2%) did not adhere to the guideline at both time points, Table 2. In these patients, the adherence to the 2022 strategy in 2022 was statistically significantly different from the adherence to the 2023 strategy in 2023 (p < 0.001).

Table 2 Adherence to GOLD Strategy in 2022 and 2023

The vast majority of patients who had prescriptions in line with the GOLD strategy at both time points, as well as those who were deviating from GOLD strategies at both time points, did not have their medication adjusted during the study period: 447 out of 454 (98.5%) and 294 out of 357 (82.4%) respectively, Figure 3A and D.

Figure 3 Treatment patterns between 2022 and 2023. This figure illustrates the medication flows and treatment patterns between 2022 and 2023 among patients who were alive at both time points (n=1074): (A) Patients treated according to GOLD at both time points (n=454, 42.3%); (B) patients treated according to GOLD in 2022 but not in 2023 (n=102, 9.5%); (C) patients not treated according to GOLD in 2022 but are treated according to GOLD in 2023 (n=161, 15.0%); (D) patients not treated according to GOLD at either time point (n=357, 34.3%). Abbreviations: LAMA, Long-Acting Muscarinic Antagonist; LABA, Long-Acting Beta-Agonist; ICS, Inhaled corticosteroids.

Of the 102 patients whose treatment went from adherent to the 2022 strategy to non-adherent to the 2023 strategy, approximately half had no change in medication (n=52, 51.0%) and half had a change in medication (n=50, 49.0%), Figure 3B. For patients whose medication remained the same, changes should have been made according to the 2023 strategy due to the use of LABA/ICS (n=22) or the use of triple therapy due to high symptom burden, which is no longer a justification for ICS in the 2023 strategy (n=5). The treatment of the remaining patients (n=25) no longer met the GOLD recommendations due to differences in exacerbations, hospitalizations, eosinophil counts, or CCQ scores between the 2022 and 2023 assessments. For example, a patient on LABA/LAMA with exacerbations had no indication for ICS in 2022 due to low eosinophil levels at the time, but became eligible for ICS in 2023 when higher eosinophil counts were measured. The 50 patients whose treatment went from adherent in 2022 to nonadherent in 2023 by changes in their medication, showed the following reasons for nonadherence to the guideline: de-escalating triple therapy to LABA/ICS (n=25), withdrawal of ICS despite exacerbations or hospitalizations (n=15) or in the presence of asthma as comorbidity (n=2), start of ICS treatment despite lack of exacerbations (n=5) or in case of eosinophils <0.1 x10⁹/L (n=1), and escalating from LABA/ICS to triple therapy in absence of exacerbations (n=2).

One hundred sixty-one patients were not treated according to the GOLD 2022 strategy, but had treatment in line with the 2023 strategy. Of these, 50 (31.1%) had their medication adjusted, so actively conforming to the strategy, and the remaining 111 patients (68.9%) did not undergo any changes in their medication regimen, Figure 3C. The majority (n=76) of the treatments of these 111 patients were now adherent to the GOLD strategy due to the differences in the recommendations between the 2022 and 2023 strategy, which aligned their existing medications with the updated recommendations. Specifically, for 59 patients on triple therapy, GOLD 2022 recommended LABA/ICS but not triple therapy, whereas the GOLD 2023 now supports the triple therapy. Similarly, 17 patients previously on LABA or LABA/LAMA had an indication for ICS in 2022, but no longer required ICS following the GOLD 2023 revision. For the remaining 35 of the 111 patients whose medication was not changed, differences in exacerbations, hospitalizations, eosinophils, or CCQ scores between the 2022 and 2023 assessments led to a new recommendation in 2023, which coincidentally aligned with their existing medication regimen. For example, one patient experienced exacerbations in 2023, which justified the triple therapy he was already receiving.

Extra Nuance of the 2025 Strategy: Clinically Stable Patients on LABA/ICS

Five of the 120 patients still on LABA/ICS in 2023 had no exacerbations, a low symptom load, and no features of asthma (4.2%), meaning that their medication regimen would have aligned with the recommendations if this current nuance of the 2025 strategy would have been valid at that time.

Differences Between Specific Patient Groups

Certain patient groups were more often received inhaled medication not in accordance with the 2023 strategy, Figure 4. When categorized by eosinophil levels into low (<0.1 x10⁹/L), medium (0.1–0.3 x10⁹/L), and high (≥0.3 x10⁹/L), there were 214 (19.9%), 513 (47.8%), and 347 (32.3%) patients in each subgroup, respectively. In the subgroup with low eosinophils, 69.6% (95% CI 63.2–75.4%) of the patients were not treated according to the 2023 strategy, compared to 40.2% (95% CI 36.0–44.5%) in the medium, and 30.0% (95% CI interval 25.4–35.0%) in the high eosinophil group (p<0.001). Patients who were treated according to the dyspnoea algorithm (n=288) were more likely to receive treatment not in line with the 2023 recommendations compared to patients who were treated according to the exacerbation algorithm (n=786): 68.4% (95% CI interval 62.8–73.5%) and 33.3% (95% CI interval 30.1–36.7%) were treated incorrectly, respectively (p<0.001). Furthermore, nonadherence to the GOLD 2023 strategy was higher in patients with less severe GOLD classifications: 56.6% (95% CI interval 46.7–65.9%) in GOLD I, 43.6% (95% CI interval 38.9–48.5%) in GOLD II, 42.8% (95% CI interval 38.2–47.5%) in GOLD III, and 28.1% (95% CI interval 20.6–36.9%) in GOLD IV subgroup (p<0.001). Among triple therapy users (n=769), 358 (46.6%) used a single device (SITT), while 411 (53.4%) used multiple devices (MITT). Nonadherence to the 2023 strategy was similar: 31.6% (95% CI interval 27.0–36.6%) in SITT compared to 29.2% (95% CI interval 25.0–33.8%) in MITT (p=0.476).

Figure 4 Percentages of patients in different subgroups who are not treated according to the 2023 strategy. Patients are categorized four times: on the far left by eosinophil levels: low (<0.1 x109/L), medium (0.1–0.3 x109/L), and high (≥0.3 x109/L); second from the left those receiving treatment based on the dyspnoea algorithm and those treated according to the exacerbation algorithm in the GOLD strategy; second from the right by GOLD classification based on airflow obstruction severity; and on the far right, triple therapy users are divided into single-inhaler triple therapy (SITT) and multiple-inhaler triple therapy (MITT) users. The blue bars represent the proportion of ICS users, while the grey bars display those without ICS. The blue dotted lines shows the percentage (42.7%) of patients that is not treated according to the GOLD report in the total cohort (n=1074). *= statistically significant difference.

In-Depth Evaluation of ICS Use

The cross-sectional analysis in 2023 showed that, among the patients alive in 2023 (n=1074), 898 patients (83.6%) were using an ICS. Of the total cohort, 625 patients (58.2%) used an ICS appropriately, while 125 patients (11.6%) correctly did not receive ICS treatment. Overtreatment was more prevalent than undertreatment: 273 patients (25.4% with 95% CI interval 22.9–28.1%) received ICS inappropriately, whereas 51 patients (4.7% with 95% CI interval 3.6–6.2%) did not receive and ICS despite having an indication according to the 2023 strategy. The highest ever recorded eosinophil value was used for the primary assessments. When instead only the most recent eosinophil count was considered, the proportion of patients classified as appropriate ICS users decreased from 58.2% to 43.8%, Figure 5A. A similar trend was seen when different timeframes for exacerbations and hospitalizations were applied: 48.4% when using data from the past year, 58.2% for the past three years, and 61.9% for the past five years, Figure 5B. Of the 898 ICS users in 2023, 59 patients (6.6%) had no exacerbations or hospitalizations in the past five years and consistently had an eosinophil count of less than 0.3.

Figure 5 Assessment of ICS treatment using different time intervals. (A) displays the proportion of patients with correct ICS use (dark green), correct ICS absence (light green), incorrect ICS use (dark Orange), and incorrect ICS absence (light orange) across different time intervals of eosinophil count. (B) presents these proportions based on various time intervals for exacerbations and hospitalizations. The black box highlights the values used for the primary assessment of inhalation treatment in this study.

When applying the 2022 and 2023 strategy at the same time point, which was the first measurement time point just before the 2022 strategy was replaced by the 2023 strategy, 1099 patients (83.4%) of the total study cohort (n=1318) were using an ICS. Among them, 88 patients had an indication for ICS at that moment according to the 2022 strategy, but would no longer have an indication once the 2023 strategy became applicable, Table 3. At the second measurement time point of this study, which occurred one year after the implementation of the 2023 strategy, 24 of these 88 patients had died. Of the 64 patients who were still alive, ICS was continued in 55 patients despite still having no indication according to the 2023 strategy, ICS was continued in five patients who had acquired an indication during that year (eg due to an exacerbation or hospitalisation), and it was appropriately discontinued in four patients.

Table 3 Longitudinal Comparison in ICS Users (n=1099) from 2022 to 2023

Discussion

In a real-world Dutch secondary care setting, this study demonstrated that only a small majority of patients were treated according to the GOLD recommendations, with a slight increase from 51.6% in 2022 to 57.3% in 2023. Some patient subgroups were at higher risk of receiving treatment that did not align with the GOLD strategy: patients with low eosinophil counts, those without exacerbations, and those with a less severe GOLD classification. Our in-depth evaluation of ICS revealed that overtreatment with ICS was common and more prevalent than undertreatment.

The adherence rate of 57% is not particularly high. One previous study assessed adherence to the GOLD strategy since the introduction of the updated treatment algorithms in the 2023 report. This cohort study of 3477 patients from 54 hospitals in South Korea reported lower adherence rates than our findings: 31.3% for the GOLD 2019 strategy and 28.0% for the GOLD 2023 strategy. Interestingly, clinicians’ adherence to the 2018 Korean national guideline was higher and more comparable (56.9%), indicating that it likely plays a more prominent role in guiding clinical practice in South Korea than the GOLD strategy.¹⁵ Other studies investigating adherence to earlier versions of the GOLD strategy reported adherence rates ranging from 36% to 60.5%.^16–19 Consistent with our findings, several studies have shown that guideline nonadherence by health care providers is more prevalent in lower-risk groups and in patients without exacerbations, in both primary and secondary care settings.^17–21 Additionally, similar to our results, overtreatment with ICS was frequently reported.^19–22

It is important to consider the barriers to implementing the GOLD recommendations in clinical practice. For example, clinicians could be reluctant to change a winning combination when patients are clinically stable Patients may also have strong preferences regarding specific inhaled medications. Furthermore, as will be discussed in more detail later in this discussion, the criteria for stopping ICS treatment are unclear. The annual release of new GOLD strategies could pose a challenge for clinicians to stay up to date if recommendations change too frequently. Davis et al investigated the knowledge and application of the GOLD recommendations through a survey among primary care and secondary care physicians in 12 countries. Primary care physicians were less familiar with the GOLD strategies than respiratory specialists (58% vs 93%). Surprisingly, the respiratory physicians did not demonstrate more guideline-concordant prescribing practices for inhaled medication than the primary care physicians, suggesting that awareness does not necessarily result in greater adherence to guidelines.¹⁰ Physicians may have followed national guidelines, which can differ from the GOLD recommendations.¹⁵ For instance, a COPD questionnaire survey in Germany showed that most (51.4%) pulmonary specialists preferred national guidelines compared to the GOLD guidelines (40.2%).²³ It could be that not all physicians are always completely at ease with the GOLD strategy. After all, the complete strategy has been compiled by experts, but changes frequently, and, to our knowledge, has never been compared to other possible strategies.

Although we observed that some patients’ inhaled medication had been appropriately adjusted in line with current guidelines, many patients without exacerbations or with low eosinophil counts were still receiving triple therapy, and others remained on LABA/ICS. The 2025 strategy offers additional nuance for the latter group, recommending the continuation of LABA/ICS for clinically stable patients without asthma, no exacerbations, and low symptom burden. However, only a small percentage (4.2%) of LABA/ICS users in our study cohort would have received the correct medication regimen if the 2025 nuances had been in effect at the time of the 2023 assessment. Adherence to the strategy changed between 2022 and 2023 in 263 patients, with some treatments going from adhering to the 2022 strategy to nonadhering to the 2023 strategy (n=102) and vice versa (n=161). Although we found that the guideline adherence rate improved slightly from 51.6% in 2022 to 57.3% in 2023, some of this improvement may be due to other factors than better adherence to the strategy per se. Not all changes in adherence could be attributed to intentional medication adjustments in response to updated recommendations, nor to the revised guidelines (in)validating existing treatment regimens. Furthermore, in a substantial proportion (n=60), changes between adherence to the 2022 and 2023 strategy were driven by variations in key clinical variables between the two assessment periods. For instance, a hospitalisation in 2019 could justify the prescription of an ICS in the 2022 assessment, as it occurred within the previous three years. However, since this hospitalisation occurred more than three years prior to the 2023 assessment, it no longer justified ICS use according to the 2023 GOLD strategy. The importance of the specific timeframes used for assessing appropriate medication was also demonstrated in our detailed assessment of ICS use within the study population.

The vast majority of our study population (83.6%) was treated with an ICS in 2023, with a substantial proportion seemingly being overtreated with ICS (25.4%), putting them at risk of side effects while probably having minimal or no benefit from the ICS. The GOLD strategies provide treatment algorithms and criteria for initiating ICS treatment, either as initial treatment or when escalation is required due to persistent exacerbations.^5,6 However, it is not specified in the GOLD strategy which eosinophil count to use for determination of treatment (eg most recent or highest value known), despite previous studies having demonstrated variability in eosinophil counts over time within individual COPD patients.^24–26 Furthermore, the GOLD strategies do not specify the timeframe for patients who are currently exacerbation-free on ICS treatment. In other words: after how many years without exacerbations is ICS treatment still justified? This absence of clear guidance regarding which eosinophil count and which timeframe for exacerbations should be applied for ICS indication is a key limitation of the current GOLD strategy. As shown by our study, increasing the timeframes for eosinophils and exacerbations increases the proportion of patients with an indication for ICS. It is possible that patients classified in our study as overtreated with ICS due to lack of recent exacerbations, originally had valid reasons for initiating ICS therapy, such as a history of multiple exacerbations at that time.

Currently, there is no conclusive evidence, and consequently no consensus, on how to manage ICS treatment once a patient becomes stable It is unclear after how many years of clinical stability without exacerbations ICS withdrawal should be considered, and whether this even should be considered if the initial indication for ICS was valid. Studies investigating the withdrawal of ICS in patients with COPD showed contradicting results.²⁷ Some studies showed that stopping ICS abruptly resulted in more exacerbations, a decline in lung function, and reduced quality of life. This was seen both in studies that compared ICS continuations to placebo,^27–29 as in one study that continued long-acting bronchodilators in both groups.³⁰ In contrast, other studies that continued long-acting bronchodilators reported no statistically significant differences in exacerbation rates between the ICS continuation and withdrawal group.^31–34 However, these results should be interpreted with caution due to the limited follow-up periods of 3 to 6 months,^31–34 observed trends toward increased exacerbations in patients who discontinued ICS,³⁴ and a higher risk of exacerbations in patients with elevated eosinophil counts.³³ The WISDOM trial is the largest trial to date investigating ICS withdrawal in COPD patients with prior exacerbations.³⁵ In contrast to the previous mentioned trials, this study first tapered the dose of ICS, rather than stopping abruptly. This approach aligns with the GINA guideline recommendations for ICS treatment in stable patients with asthma.³⁶ The WISDOM trial found no significant increase in exacerbations in the ICS withdrawal group (hazard ratio (HR) 1.06; 95% CI: 0.94–1.19). Post-hoc analysis of triple therapy users prior to enrolment showed similar results. However, the follow-up of less than one year and the observed trends towards more severe exacerbations (HR 1.20; 95% CI: 0.98–1.48), greater FEV₁ decline (−38 mL), and worse St. George’s Respiratory Questionnaire (SGRQ) scores (+0.55 vs −0.42) in the withdrawal group raise potential concerns.

The GOLD strategy lacks a recommendation regarding the withdrawal of ICS in clinically stable patients with COPD. GOLD only recommends considering ICS de-escalation in cases of side effects such as pneumonia, with the sidenote that discontinuing ICS in patients with eosinophil counts above 0.3 x10⁹/L could lead to more exacerbations.¹ Other guidelines provide more specific guidance on when ICS could be discontinued if patients remain free of exacerbations. For instance, the most recent Dutch COPD guidelines for both primary and secondary care suggest considering abrupt withdrawal of ICS after two years without exacerbations,^2,3 even though it is acknowledged in the latter that this recommendation is not supported by any evidence.³ Since one of the primary reasons for initiating ICS is to reduce exacerbations, stopping treatment when it has been successful (ie no exacerbations) seems contradictory. In our study, only a small proportion of patients (6.6%) consistently had eosinophil counts below 0.3 x10⁹/L and remained free of exacerbations and hospitalisations for at least five years. Furthermore, it should be noted that a recent systematic review and meta-analysis showed a reduction in all-cause mortality in COPD patients with ICS treatment.³⁷ The effect of discontinuing ICS on this mortality benefit is unknown. The lack of evidence on ICS withdrawal in stable patients may make clinicians reluctant to stop ICS therapy. This is reflected in our study by the small proportion of ICS users whose indication under the 2022 strategy no longer applied in 2023, and who actually had their ICS discontinued (4 out of 59). Conducting a prospective study with long-term follow-up to evaluate the impact of ICS discontinuation in clinically stable COPD patients could help fill this evidence gap.

Our study has some limitations. The data were extracted from a single secondary care centre in the Netherlands, yet including multiple prescribers. As shown in the baseline characteristics, our study population includes a relatively high proportion of triple therapy users, suggesting that our hospital may have a more severe COPD population, as indicated by the mortality rate. Although the hospital serves as a regional centre of expertise due to its coordination of care, its COPD population is comparable to those of general hospitals. In the Netherlands, COPD care is guided by either the GOLD strategy or the national guideline, which largely align with GOLD recommendations on pharmacological treatment.^2,3 Therefore, we expect that the findings of this study are also applicable to other countries where COPD care is guided by the GOLD strategy. This is supported by the similarity of our results to those of previous studies in a secondary care setting in other countries.^16–20,22 Consistent with previous studies, our study showed that clinicians were less likely to adhere to the GOLD recommendations regarding inhaled medication when treating low-risk patients.^17–21 Moreover, Rodrigues et al showed in a Dutch primary care setting that 36.4% of the COPD patients did not receive any inhaled medication within the first three months after diagnosis.³⁸ As most COPD patients in the Netherlands are treated in primary care, overall adherence to the GOLD strategy may be even lower in the general population than in our study cohort. Furthermore, patients with missing CCQ or eosinophil data were excluded from the study cohort. This may have introduced some selection bias towards patients with more severe COPD, since these values are more likely to be missing in patients with stable, less severe disease. This is also reflected in treatment patterns, with the excluded cohort receiving triple therapy less frequently. One might argue that in most patients, according to the latest recommendations by GOLD, it is not possible to determine whether treatment is appropriate without knowing the eosinophil count and CCQ score or another disease burden questionnaire, such as the Medical Research Council (MRC) or COPD Assessment Test (CAT). Physicians should be encouraged to assess the eosinophil count and disease burden in all COPD patients, including those with less severe and more clinically stable disease, to determine appropriate treatment. Due to the design of the study, it was not possible to account for the potential use of systemic corticosteroids prior to eosinophil testing. Similarly, appropriate de-escalation of ICS treatment based on the presence of pneumonia or other side effects could not be assessed. Given that, according to our assessment, only a small number of patients were undertreated, we expect that this did not occur often. Also, prednis(ol)one prescriptions were used as a surrogate for exacerbations. Exacerbations could have been missed if pharmacy records were not available in the “Landelijk Schakelpunt (LSP)” or if exacerbations were not treated with prednis(ol)one. However, given that most pharmacies participate in the LSP (95%) and most patients granted consent (90%), we assume that the number of prednisone prescriptions not recorded in our dataset is limited.¹² Furthermore, based on our clinical experience and in line with GOLD and national guidelines, we anticipate that it is very uncommon for patients with COPD exacerbations to be treated with antibiotics without the use of prednisone. Due to the limitations of this retrospective study, our database did not contain all factors that could influence adherence to the guidelines at patient-, prescriber- or time-related levels. These factors, for instance, include patient preferences, side-effects, and previous ICS response. As a next step following this descriptive study, a prospective study incorporating these factors in a multivariate regression analysis or an in depth qualitative study would help clarify what drives guideline non-adherence.

Conclusion

A substantial proportion of patients with COPD are not treated in accordance with prevailing recommendations in the GOLD strategy, particularly those with a low eosinophil count, those without exacerbations or hospitalisations, and those with a less severe GOLD classification. This study demonstrates that even after the 2023 updates to inhaled medication recommendations, guideline adherence improved only slightly and few had their medication actively adjusted, highlighting the real-world challenges of implementing guideline recommendations in daily practice. While there may be valid reasons for deviating from the guidelines, these findings should prompt treating physicians to reconsider the appropriateness of the current treatment for their patients. ICS overtreatment was frequent, posing an increased risk of unwanted short and long term side effects. This study highlighted key limitations of the current GOLD strategy, including the lack of consensus on whether to continue ICS treatment in stable COPD patients who initially had an appropriate indication, and the absence of clear guidance on which eosinophil count and which exacerbation timeframe should be use to guide ICS continuation. Conducting additional consensus studies on ICS tapering could provide clearer guidance for clinicians on which patients can safely discontinue ICS and which cannot, potentially leading to greater adherence to guidelines in clinical practice.

Data Sharing Statement

All data relevant to the study are included in the article or uploaded as supplementary information. The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Ethics Approval and Consent to Participate

Ethics approval for this study was waived by the Institutional Research Board of the Franciscus Gasthuis & Vlietland, Rotterdam, the Netherlands (identification number 2023-037). This research was declared outside the scope of the Medical Research Involving Human Subjects Act, as it used routinely collected healthcare data that had been pseudonymized. The Institutional Research Board of Franciscus Gasthuis & Vlietland waived the requirement to obtain informed consent, because the data is collected in a coded (pseudonymized) form, and obtaining consent is not feasible due to the large number of patients.

Acknowledgments

We are grateful to dr. Simone Rauh for her help in processing the data.

Author Contributions

LC and JiV initiated this study. 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

The faculty received an unrestricted grant from Chiesi Pharmaceuticals B.V. Furthermore, the funding includes contributions from the research group in the Franciscus Gasthuis & Vlietland hospital and transformation funds from health insurance companies.

Disclosure

LC: the faculty has received an unrestricted grant from Chiesi Pharmaceuticals B.V (outside the scope of this work). JvB: has received grants and/or consultancy fees from ALK-Abello, AstraZeneca, Chiesi, European Commission COST (COST Action 19132), GSK, Novartis, Pfizer, Teva, Trudell Medical, and Vertex, outside the submitted work and all paid to his institution (UMCG). HK: has received research/educational grants and served on advisory boards for Boehringer Ingelheim, GSK and Novartis, and has served on advisory boards for AstraZeneca and Chiesi. JiV: reports unrestricted faculty research grants from GSK, Teva, AZ, Chiesi, Sanofi, and speaker fees from AZ, GSK, Sanofi, Chiesi, Stichting RoLeX and Health Investment. The authors report no other conflicts of interest in this work.

References

1. Global Strategy for the Diagnosis, Management, and Prevention of Chronic Obstructive Pulmonary Disease (2025 report). Global Initiative for Chronic Obstructive Lung Disease.

2. Bischoff EBM, Broekhuizen L, Donkers J, et al. 1NHG-standaard COPD. Nederlands Huisartsen Genootschap. 2021. Available from: https://richtlijnen.nhg.org/standaarden/copd. Accessed November 20, 2025.

3. Federatie Medisch Specialisten (FMS) richtlijn COPD. Nederlandse Verenigingvan Artsen voor Longziektenen Tuberculose (NVALT). Available from: https://richtlijnendatabase.nl/richtlijn/copd/startpagina_-_copd.html. Accessed November 20, 2025.; 2023.

4. Agusti A, Celli BR, Criner GJ, et al. Global Initiative for chronic obstructive lung disease 2023 report: GOLD executive summary. Eur Respir J. 2023;61(4):2300239. doi:10.1183/13993003.00239-2023

5. Global strategy for prevention, diagnosis and management of: 2023 report COPD. Available from: https://goldcopd.org/2023-gold-report-2/2023. Accessed November 20, 2025.

6. 2022 Global Strategy for Prevention, Diagnosis and Management of COPD. Available from: https://goldcopd.org/2022-gold-reports/2022. Accessed November 20, 2025.

7. Lipson DA, Barnhart F, Brealey N, et al. Once-daily single-inhaler triple versus dual therapy in patients with COPD. N Engl J Med. 2018;378(18):1671–1680. doi:10.1056/NEJMoa1713901

8. Rabe KF, Martinez FJ, Ferguson GT, et al. Triple inhaled therapy at two glucocorticoid doses in moderate-to-very-severe COPD. N Engl J Med. 2020;383(1):35–48. doi:10.1056/NEJMoa1916046

9. Quint JK, Ariel A, Barnes PJ. Rational use of inhaled corticosteroids for the treatment of COPD. NPJ Prim Care Respir Med. 2023;33(1):27. doi:10.1038/s41533-023-00347-6

10. Davis KJ, Landis SH, Oh YM, et al. Continuing to Confront COPD International Physician Survey: physician knowledge and application of COPD management guidelines in 12 countries. Int J Chron Obstruct Pulmon Dis. 2015;10:39–55. doi:10.2147/COPD.S70162

11. von Elm E, Altman DG, Egger M, et al. The strengthening the reporting of observational studies in epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol. 2008;61(4):344–349. doi:10.1016/j.jclinepi.2007.11.008

12. Samenvatting en Conclusie Apotheken, RIVM Onderzoek ICT in de Zorg. Rijksinstituut voor Volksgezondheid en Milieu (RIVM). 2016. Available from: https://www.rivm.nl/documenten/samenvatting-resultaten-apotheken. Accessed, 2025.

13. Sundh J, Stallberg B, Lisspers K, Kampe M, Janson C, Montgomery S. Comparison of the COPD assessment test (CAT) and the Clinical COPD questionnaire (CCQ) in a clinical population. COPD. 2016;13(1):57–65. doi:10.3109/15412555.2015.1043426

14. Smid DE, Franssen FME, Gonik M, et al. Redefining cut-points for high symptom burden of the global initiative for chronic obstructive lung disease classification in 18,577 patients with chronic obstructive pulmonary disease. J Am Med Dir Assoc. 2017;18(12):1097e11–e24. doi:10.1016/j.jamda.2017.09.003

15. Han SM, Kim HS, Park SY, et al. Adherence to pharmacological management guidelines for stable chronic obstructive lung disease. Tuberc Respir Dis. 2025;88(2):310–321. doi:10.4046/trd.2024.0130

16. Mannino DM, Yu TC, Zhou H, Higuchi K. Effects of GOLD-adherent prescribing on COPD symptom burden, exacerbations, and health care utilization in a real-world setting. Chronic Obstr Pulm Dis. 2015;2(3):223–235. doi:10.15326/jcopdf.2.3.2014.0151

17. Grewe FA, Sievi NA, Bradicich M, et al. Compliance of pharmacotherapy with GOLD guidelines: a longitudinal study in patients with COPD. Int J Chron Obstruct Pulmon Dis. 2020;15:627–635. doi:10.2147/COPD.S240444

18. Palmiotti GA, Lacedonia D, Liotino V, et al. Adherence to GOLD guidelines in real-life COPD management in the Puglia region of Italy. Int J Chron Obstruct Pulmon Dis. 2018;13:2455–2462. doi:10.2147/COPD.S157779

19. Scalone G, Nava S, Ventrella F, et al. Pharmacological approach and adherence to treatment recommendations in frequently and non-frequently exacerbating COPD patients from Italy: MISTRAL – The prospective cohort, observational study. Pulm Pharmacol Ther. 2018;53:68–77. doi:10.1016/j.pupt.2018.09.001

20. Alabi FO, Alkhateeb HA, Zibanayi MT, et al. The adherence to and utility of the Global Initiative for Chronic Obstructive Lung Disease guidelines for treating COPD among pulmonary specialists: a retrospective analysis. BMC Pulm Med. 2023;23(1):216. doi:10.1186/s12890-023-02503-7

21. Mangold V, Boesing M, Berset C, et al. Adherence to the GOLD guidelines in primary care: data from the Swiss COPD cohort. J Clin Med. 2023;12(20):6636. doi:10.3390/jcm12206636

22. Larsson K, Ekberg-Jansson A, Stridsman C, Hanno M, Vanfleteren L. Adherence to treatment recommendations for chronic obstructive pulmonary disease – results from the swedish national airway register. Int J Chron Obstruct Pulmon Dis. 2021;16:909–918. doi:10.2147/COPD.S300299

23. Glaab T, Vogelmeier C, Hellmann A, Buhl R. Guideline-based survey of outpatient COPD management by pulmonary specialists in Germany. Int J Chron Obstruct Pulmon Dis. 2012;7:101–108. doi:10.2147/COPD.S27887

24. Singh D, Kolsum U, Brightling CE, et al. Eosinophilic inflammation in COPD: prevalence and clinical characteristics. Eur Respir J. 2014;44(6):1697–1700. doi:10.1183/09031936.00162414

25. Kwok WC, Chau CH, Tam TCC, Lam FM, Ho JCM. Variability of blood eosinophil count at stable-state in predicting exacerbation risk of chronic obstructive pulmonary disease. Int J Chron Obstruct Pulmon Dis. 2023;18:1145–1153. doi:10.2147/COPD.S401357

26. Bafadhel M, Pavord ID, Russell REK. Eosinophils in COPD: just another biomarker? Lancet Respir Med. 2017;5(9):747–759. doi:10.1016/S2213-2600(17)30217-5

27. Georgiou A, Ramesh R, Schofield P, White P, Harries TH. Withdrawal of inhaled corticosteroids from patients with COPD; effect on exacerbation frequency and lung function: a systematic review. Int J Chron Obstruct Pulmon Dis. 2024;19:1403–1419. doi:10.2147/COPD.S436525

28. van der Valk P, Monninkhof E, van der Palen J, Zielhuis G, van Herwaarden C. Effect of discontinuation of inhaled corticosteroids in patients with chronic obstructive pulmonary disease: the COPE study. Am J Respir Crit Care Med. 2002;166(10):1358–1363. doi:10.1164/rccm.200206-512OC

29. Choudhury AB, Dawson CM, Kilvington HE, et al. Withdrawal of inhaled corticosteroids in people with COPD in primary care: a randomised controlled trial. Respir Res. 2007;8(1):93. doi:10.1186/1465-9921-8-93

30. Wouters EF, Postma DS, Fokkens B, et al. Withdrawal of fluticasone propionate from combined salmeterol/fluticasone treatment in patients with COPD causes immediate and sustained disease deterioration: a randomised controlled trial. Thorax. 2005;60(6):480–487. doi:10.1136/thx.2004.034280

31. Rossi A, van der Molen T, Del Olmo R, et al. INSTEAD: a randomised switch trial of indacaterol versus salmeterol/fluticasone in moderate COPD. Eur Respir J. 2014;44(6):1548–1556. doi:10.1183/09031936.00126814

32. Frith PA, Ashmawi S, Krishnamurthy S, et al. Efficacy and safety of the direct switch to indacaterol/glycopyrronium from salmeterol/fluticasone in non-frequently exacerbating COPD patients: the FLASH randomized controlled trial. Respirology. 2018;23(12):1152–1159. doi:10.1111/resp.13374

33. Chapman KR, Hurst JR, Frent SM, et al. Long-term triple therapy de-escalation to indacaterol/glycopyrronium in patients with chronic obstructive pulmonary disease (SUNSET): a randomized, double-blind, triple-dummy clinical trial. Am J Respir Crit Care Med. 2018;198(3):329–339. doi:10.1164/rccm.201803-0405OC

34. Harries TH, Gilworth G, Corrigan CJ, et al. Withdrawal of inhaled corticosteroids from patients with COPD with mild or moderate airflow limitation in primary care: a feasibility randomised trial. BMJ Open Respir Res. 2022;9(1). doi:10.1136/bmjresp-2022-001311.

35. Magnussen H, Disse B, Rodriguez-Roisin R, et al. Withdrawal of inhaled glucocorticoids and exacerbations of COPD. N Engl J Med. 2014;371(14):1285–1294. doi:10.1056/NEJMoa1407154

36. Global Strategy for Asthma Management and Prevention. Global Initiative for Asthma; 2024. Available from: https://ginasthma.org/2024-report/. Accessed November 20, 2025.

37. Rogliani P, Manzetti GM, Gholamalishahi S, Bafadhel M, Calzetta L. Inhaled corticosteroids in chronic obstructive pulmonary disease: a systematic review and meta-analysis on mortality protection – making a long story short. Expert Rev Respir Med. 2025;1–11. doi:10.1080/17476348.2025.2465853

38. Rodrigues G, Antao J, Deng Q, et al. Trends in initial pharmacological COPD treatment in primary care (2010-2021): a population-based study using the PHARMO Data Network. Respir Res. 2024;25(1):447. doi:10.1186/s12931-024-03073-w

Introduction

Inflammatory bowel diseases (IBD), including Crohn’s disease (CD) and ulcerative colitis (UC), result from interactions between the host, environmental factors, and gut microbiome.^1,2 A common feature of IBD is microbial community alteration, characterized by reduced microbial diversity, depletion of short-chain fatty acid–producing bacteria, enrichment of opportunistic taxa, and disrupted metabolic pathways that change with disease activity and therapeutic interventions.^3–5 Recent longitudinal and multi-omics studies have shown that these changes in compositional and functional shifts are closely related to host transcriptomic, proteomic, and metabolomic signatures, underscoring the centrality of host–microbe interactions in disease pathogenesis.^6–8

Despite these advances, significant clinical unmet needs remain. The diagnosis of IBD is often delayed because symptoms overlap with those of other gastrointestinal disorders,^9,10 and current biomarkers such as C-reactive protein and fecal calprotectin lack disease specificity and prognostic utility.^11–13 Reliable microbial or molecular biomarkers capable of predicting treatment response or relapse risk are still lacking, highlighting the need for integrative analytic frameworks that can bridge the gap between discovery and clinical application.^14,15 Traditional statistical analyses, including univariate (Linear discriminant analysis Effect Size, LEfSe) or multivariate regression models (Microbiome Multivariable Association with Linear Models, MaAsLin2), have been instrumental in identifying differentially abundant taxa and host factors.^16,17 However, these approaches are limited in capturing nonlinear interactions and complex dependencies within high-dimensional, multi-omics datasets.^18,19

These datasets are intrinsically high-dimensional, sparse, and heterogeneous across cohorts, sampling sites, sequencing platforms, and geographies, with outcomes that can vary considerably depending on the analytical methods applied. Longitudinal and cross-cohort studies have revealed significant inter-individual variability and nonlinear dependencies among microbes, host factors, and metabolites, which makes finding biomarkers more complicated.^7,20 Additionally, large-scale integration of metagenomic and metabolomic data has shown both the potential to identify disease-related features and the ongoing challenge of aligning results across studies. This emphasizes the need for strong computational methods to handle this complexity.^6,21

In this context, machine learning (ML) has emerged as a robust framework for characterizing gut microbial changes in IBD.²² ML approaches are particularly well-suited for (i) extracting predictive patterns from high-dimensional data, (ii) integrating heterogeneous inputs spanning microbial taxa, functional genes, host omics, and clinical metadata, and (iii) delivering measurable classification performance for diagnostic, prognostic, and therapeutic stratification.²³ Applications in IBD cohorts have demonstrated that ML models outperform traditional statistical methods in discriminating cases from controls, differentiating subtypes, and predicting treatment responses.^4,24 Simultaneously, benchmarking studies and recent state-of-the-art reviews have emphasized critical challenges—including external validation, model calibration, interpretability, and transparent reporting—that must be addressed before ML-driven microbiome biomarkers can be reliably translated into clinical settings.^18,23,25

Building on recent reviews that highlighted the importance of methodological robustness, reproducibility, and the need for clinical translation in AI-driven microbiome research,^26,27 this review extends these perspectives by incorporating frameworks such as time-series and causal inference modeling, multi-omics data integration, and foundation-model approaches to enhance translational relevance. We provide a comprehensive overview of recent applications of ML for characterizing gut microbiome alterations in IBD, emphasizing commonly used algorithms, analytical pipelines, and model evaluation strategies. Finally, we highlight the diagnostic and prognostic potential of ML-based approaches and discuss their strengths, limitations, and key considerations in translating microbiome-derived insights into clinically applicable tools for precision medicine.

Microbiome Dysbiosis in IBD: From Traditional Studies to Multi-Omics Analyses

Early studies on gut microbiome in IBD primarily relied on 16S rRNA gene sequencing, which enabled cost-effective profiling of microbial communities at the genus or family level. These studies provided the first evidence of disease-associated dysbiosis characterized by reduced alpha diversity, enrichment of pathobionts, and depletion of commensals.^3,5 Initial analytical approaches focused mainly on univariate comparisons of relative abundances, typically using methods such as LEfSe.¹⁷ Ma et al applied LEfSe analysis to compare patients with early-stage CD, advanced CD, and healthy controls, identifying enrichment of Parabacteroides and Lachnospiraceae incertae sedis in early CD, expansion of Escherichia-Shigella and Proteus in advanced CD, and preservation of short-chain fatty acid-producing taxa such as Roseburia and Butyricicoccus in healthy controls.²⁸ These findings demonstrate how LEfSe highlights disease-stage-specific microbial signatures while underscoring its reliance on univariate contrasts.

However, subsequent microbiome analyses have increasingly adopted multivariate analyses that incorporated metadata-based adjustments and advanced statistical modeling. Tools, such as MaAsLin2¹⁶/MaAsLin3²⁹ and ANCOM-BC³⁰/ANCOM-BC2,³¹ further enhance feature selection by integrating clinical metadata and complex covariates. Chen et al used MaAsLin2 to identify taxa whose relative abundances differed in quiescent CD (CD-R), independent of active inflammation and environmental or genetic confounders.³² They found that Faecalibacterium, Dorea, and Fusicatenibacter were significantly decreased in patients with CD-R compared to healthy first-degree relatives and non-relative healthy controls. Rosso et al applied ANCOM-BC to fecal bacterial profiles of patients with IBD (both UC and CD) vs non-IBD controls in Buenos Aires.³³ Among the bacterial genera that were differentially abundant in ANCOM-BC, Bifidobacterium was enriched in patients with UC compared to non-IBD controls, whereas in patients with CD, Bifidobacterium, Bacteroides, Lactobacillus, and Faecalibacterium were identified as differentially abundant taxa. Together, these approaches reduced the analytical bias and enabled a more reproducible investigation of disease-specific taxa, marking a substantial advancement from exploratory compositional analyses to clinically grounded biomarker discovery. In addition, other methods such as ALDEx2³⁴ and LinDA,³⁵ although not discussed in detail in this review, represent parallel efforts to further mitigate the biases inherent to differential abundance analysis.

The advent of high-throughput shotgun metagenomic sequencing has provided a technological leap, offering strain-level resolution and direct access to microbial functional repertoires, including genes and pathways relevant to host metabolism and immunity.³⁶ In a recent multibiome study, Akiyama et al applied shotgun metagenomic sequencing with MaAsLin2 across Japanese and external validation cohorts to characterize microbial signatures in IBD.³⁷ Both UC and CD are associated with the depletion of short-chain fatty acid-producing bacteria and the enrichment of Enterococcus faecium and Bifidobacterium species. Notably, Escherichia coli specifically increased only in patients with CD, highlighting its potential role as a disease-specific microbial marker.

Importantly, gut microbiome dysbiosis in patients with IBD is not limited to bacteria. Studies on the gut virome^38,39 have reported an increased abundance of bacteriophages (eg, Caudovirales) and decreased abundance of eukaryotic viruses, while mycobiome studies have identified altered fungal-bacterial interactions and increased abundance of Candida species.⁴⁰ Beyond the microbial taxa themselves, metabolomic and proteomic analyses revealed disease-associated alterations in bile acids, short-chain fatty acids, and amino acid derivatives, providing functional readouts that link microbial activity to host immune and epithelial responses.^41,42 Together, these multi-omics approaches provide a more comprehensive understanding of microbial changes and its systemic effects.

However, the integration of multiple data layers has resulted in significant analytical challenges. Microbiome and multi-omics datasets are intrinsically high-dimensional, sparse, and heterogeneous with marked variability across cohorts, sequencing platforms, and sample types.^18,19,43 Traditional univariate and linear statistical approaches are often insufficient for capturing nonlinear dependencies and cross-modal interactions. Although analytical tools such as LEfSe and MaAsLin2 are widely used for differential abundance and association testing, they are inherently sensitive to compositional bias, differences in sequencing depth, and batch effects.^19,44 LEfSe operates on relative abundances and may generate spurious associations when sample library sizes vary, whereas MaAsLin2 assumes linear relationships and is affected by zero inflation and normalization artifacts. These issues have been repeatedly highlighted in recent benchmarking studies, emphasizing the need for compositional data–aware methods and rigorous normalization to ensure cross-cohort reproducibility and biological validity.⁴⁵ Collectively, these challenges have motivated the adoption of ML and related artificial intelligence (AI) approaches, which are particularly well-suited for recognizing complex patterns in high-dimensional data and enabling multimodal integration.^46,47 Leveraging these advanced computational frameworks, researchers aim to move beyond traditional analyses toward more reproducible, predictive, and clinically interpretable models of gut microbiome dysbiosis in IBD.

Machine Learning Applications in IBD Microbiome Research

As the complexity of microbiome data in IBD has become more apparent, ML approaches have been increasingly adopted to address the limitations of conventional statistical analyses.^18,48 Unlike traditional differential abundance methods, ML algorithms can handle high-dimensional and heterogeneous data and discover subtle nonlinear patterns that may better capture disease-associated microbial biomarkers.^49,50 These shifts have established ML as a key tool for developing diagnostic and prognostic models of IBD.^51,52

Disease Classification and Subtype Diagnosis

One of the earliest and most widely explored applications of ML in microbiome research was disease diagnosis. Classifier models trained on microbial abundance profiles or functional gene features have been used to discriminate patients with IBD from healthy controls,²⁴ distinguish patients with CD from those with UC,⁵⁰ and predict clinical outcomes such as flare risk, therapeutic response, or remission status.^52,53 In many cases, ML classifiers have achieved significant performance as measured via receiver operating characteristic (ROC) curves, the area under the curves (AUCs), sensitivity, and specificity, often surpassing the predictive capacity of conventional biomarkers such as fecal calprotectin.²⁴ For example, Liñares-Blanco et al applied a panel of supervised learning algorithms to fecal microbiome data and developed a robust signature capable of classifying not only IBD versus controls but also differentiating CD from UC.⁵⁰ Their random forest (RF)-based model achieved high AUC values across cross-validation, and feature selection identified a reproducible set of bacterial taxa that contributed most strongly to classification. Similarly, Park et al reported that a fecal microbiome-based ML model accurately discriminated between CD and UC, further confirming that microbial signatures can serve as reliable classifiers of IBD subtypes.⁵¹ Recent large-scale studies have underscored the diagnostic power of microbiome-informed ML. Zheng et al’s multi-ethnic cohort analysis demonstrated that an ML classifier trained on microbial taxonomic features achieved AUCs exceeding 0.7~0.9, outperforming conventional biomarkers such as fecal calprotectin in identifying both UC and CD.²⁴ Similarly, Kim et al constructed an ML model using Korean fecal metagenomes that reliably differentiated patients with IBD from healthy patients, emphasizing the generalizability of microbiome signatures across populations.⁵⁴ Together, these studies highlight that microbiome-based ML frameworks may evolve into noninvasive diagnostic tools capable of complementing or reducing the need for invasive endoscopy.

Biomarker Discovery and Algorithmic Approaches

In addition to its classification capabilities, ML plays a key role in the discovery of disease-specific features and biomarkers.^55,56 By ranking microbial taxa, genes, or metabolites according to their predictive importance, algorithms not only improve model interpretability but also generate hypotheses about how microbes might contribute to diseases. This capability has enabled the development of candidate biomarker panels that integrate microbial and metabolic features with clinical metadata, thereby providing a more comprehensive view of IBD pathophysiology. Among the various algorithms employed, RF⁵⁷ is the most widely used in microbiome studies because of its robustness to noise, sparsity tolerance, and straightforward feature importance readouts.⁵⁸ Knights et al highlighted its stability and performance compared to traditional statistical approaches.⁵⁹ Kim et al further applied RF–based feature selection to identify microbial markers associated with IBD, distinguishing taxa such as Veillonella and Escherichia-Shigella between patients with UC and those with CD.⁵⁴ Manandhar et al also applied RF classifiers trained on gut microbiome features to diagnose IBD and differentiate CD from UC using the American Gut Project data comprising 729 patients with IBD and 700 healthy patients.⁴⁷ With 50 LEfSe-selected taxa, the RF model achieved a test AUCs of > 0.80 for IBD versus controls; using the top 500 high-variance operational taxonomic units (OTUs) further improved performance to an AUCs > 0.82. For subtype classification (331 CD vs 141 UC), RF models trained on either differential taxa or high-variance OTUs attained test AUCs > 0.90, underscoring their strong discriminatory power and potential clinical utility.

Although RF remains the most commonly used approach, other algorithms, including support vector machines (SVMs) and boosting frameworks (eg, XGBoost and AdaBoost), are also widely applied.^60,61 In an extensive multiomics study of CD location subtypes, Gonzalez et al systematically evaluated ten different algorithms (including RFs, extra trees, decision trees, SVMs, multilayer perceptron classifiers, voting classifiers, naïve Bayes, k-nearest neighbors, logistic regression, and AdaBoost) for their ability to classify disease subtypes. The authors identified the best-performing models and applied them to downstream analyses, revealing distinct associations according to disease location. Specifically, colonic CD exhibited a greater similarity to ulcerative colitis and showed stronger associations with Bacteroides vulgatus and neutrophil activity, whereas ileal CD was more closely linked to bile acid metabolism and demonstrated marked alterations in Faecalibacterium prausnitzii. These findings underscore how both the model choice and feature modality substantially influence predictive performance.⁶²

Besides single-algorithm approaches, ensemble methods are also advantageous. Ha et al trained eight ML algorithms, including RF, deep neural networks (DNNs), logistic regression, k-nearest neighbors, decision trees, gradient boosting, and SVMs, on multi-cohort pediatric IBD data. An ensemble model combining three top-performing classifiers (DNN, logistic regression, and SVMs) achieved the highest accuracy in predicting future remission, outperforming any single algorithm.⁵² The characteristics of each ML/DL algorithm are summarized in Table 1.

Table 1 Comparative Characteristics and Performance Trends of ML Algorithms in IBD Microbiome Research

Challenges and Limitations of ML in IBD Microbiome Research

However, despite these advances, several challenges remain unresolved. Overfitting is a recurrent issue due to the imbalance between high-dimensional microbial features and relatively small sample sizes. A recent review by Dudek et al emphasized the risks of data leakage and overfitting in microbiome ML workflows, calling for more rigorous cross-validation and model-locking strategies.⁶⁹ In addition to overfitting, performance metrics such as AUCs or accuracy often rely solely on internal cross-validation and may overestimate true generalizability. Robust evaluation requires nested cross-validation, bootstrapping, and, ideally, independent external validation to ensure fair model assessment and prevent information leakage.^48,70 As demonstrated by Kubinski et al, model performance varies markedly across preprocessing choices and cohorts, underscoring the difficulty of achieving cross-cohort reproducibility.⁷¹ Moreover, upstream sources of technical noise—including differences in normalization procedures, zero-inflated feature distributions, and the inherent sparsity and compositionality of microbial data—can substantially distort model behavior and feature importance rankings, further complicating reproducibility and external validation.^71–73 Furthermore, as highlighted by Papoutsoglou et al, the lack of standardized preprocessing pipelines and benchmarking datasets undermines reproducibility and complicates comparisons across studies.⁴⁸ Collectively, these findings demonstrate that while ML has brought substantial progress beyond traditional statistical approaches in IBD microbiome research, persistent barriers, including overfitting, limited generalizability, and lack of methodological standardization, must be addressed. Rigorous external validation, harmonized data pipelines, and the integration of interpretable ML frameworks are essential to ensure that microbiome-based models can transition into reliable clinical applications.⁷⁴

Emerging Approaches Beyond Traditional ML

While traditional ML approaches such as RF and SVMs have shown strong utility in IBD microbiome research, recent years have seen the rise of deep learning (DL) and next-generation AI frameworks that aim to capture even more complex and nonlinear relationships within multi-omics data.^66,75 Unlike traditional methods that often rely on predefined or engineered features, DL methods use representation learning to automatically extract latent patterns from raw or minimally processed inputs.^75,76 Oh and Zhang reported DeepMicro as a notable early framework that employed autoencoders to compress high-dimensional microbiome profiles into compact latent representations, thereby enhancing the disease classification performance. Building on this foundation, subsequent DL applications have explored a range of architectures: autoencoders for dimensionality reduction and feature learning,⁷⁵ and convolutional neural networks (CNNs) to identify structured patterns in microbial abundance matrices.⁷⁷ More advanced models, such as the Multimodal Variational Information Bottleneck (MVIB), further integrate multi-omics layers and clinical metadata, highlighting how DL can support more accurate and clinically relevant predictions.⁷⁸ Collectively, these approaches emphasize the potential of DL to surpass conventional ML, offering scalable and flexible tools for predictive microbiome research.

Recently, attention has shifted toward the development of foundation models that leverage large-scale pretraining on biological sequences to generate generalized representations. The Microbial General Model (MGM) was one of the first attempts to construct a microbiome-focused foundation model trained on more than 260,000 microbiome samples using transformer-based pretraining.⁷⁹ Although still in its early stages, the MGM demonstrates how models tailored to microbial communities can eventually classify community types, identify keystone taxa, and capture longitudinal dynamics, thereby providing a scalable framework for studying microbiome-driven processes in IBD. In parallel, host-centric genome foundation models, such as AlphaGenome⁸⁰ and Nucleotide Transformer,⁸¹ have been developed to annotate genetic variants and predict their regulatory effects in human genomics. Although not originally designed for microbial applications, these host-oriented models provide transferable embeddings that may help bridge host genetic information with microbial-community dynamics. Together, microbiome- and host-focused foundation models highlight a forward-looking trajectory in which unified representational spaces may enable multi-omics integration and deepen our understanding of host–microbiome interactions in IBD.

In parallel, recent multi-omics studies have begun to adopt time-series and causal inference models that capture longitudinal microbiome–host interactions and reveal potential drivers of disease dynamics.^82,83 Dynamic Bayesian networks, Granger-causality analysis, and structural causal modeling frameworks allow researchers to infer temporal dependencies and directionality in microbial and metabolic shifts, providing insights that static cross-sectional analyses cannot capture.⁸⁴ These temporal and causal modeling approaches are crucial for understanding how microbial community states evolve during inflammation, remission, and therapeutic intervention in IBD. In addition, multi-omics data-fusion frameworks have rapidly advanced predictive microbiome analytics by integrating heterogeneous data layers within shared latent spaces. Representative examples include Multi-Omics Factor Analysis (MOFA+), which applies probabilistic factor modeling to uncover coordinated variance across omics layers;⁸⁵ DIABLO, a multivariate approach for discriminative integration of multi-omic modalities;⁸⁶ and deep representation learning frameworks, such as variational autoencoder–based or multimodal transformer architectures, which embed genomic, transcriptomic, and metabolomic profiles into unified representations for improved disease prediction and mechanistic interpretation.⁷⁵ More recently, the MintTea (Multi-omics Integration through Temporal Embedding and Attention) framework has extended this concept by integrating longitudinal multi-omics data through attention-based deep learning, enabling the identification of dynamic, disease-associated molecular modules across time and modalities.⁸⁷ These integrative frameworks exemplify the shift from isolated single-omic analysis toward data-fusion models that bridge microbial, metabolic, and host dimensions of IBD.

As these models grow in complexity, concerns regarding their interpretability have become more pressing. To ensure clinical relevance and acceptance, the integration of explainable AI (XAI)⁸⁸ methods, such as SHapley Additive exPlanation (SHAP) values, attention weight visualization, and concept bottleneck models, has been proposed.⁸⁹ For instance, Novielli et al applied SHAP values to microbiome classification tasks and showed that taxa such as Fusobacterium and Parvimonas were key drivers of colorectal cancer predictions, thereby providing transparent links between features and model outputs.⁹⁰ Similarly, Onwuka et al demonstrated that the SHAP-based prioritization of fecal and plasma metabolites could highlight IBD-associated signatures, reinforcing the role of XAI in biomarker discovery and clinician application.⁹¹ In line with these advances, a recent review by Kim et al emphasized that the field is shifting from descriptive profiling of dysbiosis toward more predictive and analytic (including prescriptive) applications, while highlighting ongoing challenges such as biological and technical heterogeneity, limited generalizability across cohorts, and the need for methodological standardization and external validation.⁹²

Clinical Translation and Challenges

The application of ML and DL to the IBD microbiome data has generated promising diagnostic models and candidate biomarker panels.^93,94 However, the path from computational discovery to clinical implementation remains complex and requires careful consideration of several limitations.^95,96 While traditional statistical analyses and ML/DL-based approaches have successfully identified bacterial taxa associated with disease states, causal or mechanistic links between these microbes and host physiology have been established only for a limited subset. Most computational analyses cannot by themselves determine whether these microbial signatures drive, result from, or merely correlate with inflammation. For example, several genera including Faecalibacterium frequently identified by differential abundant taxa analysis are known to produce short-chain fatty acids that modulate intestinal immune responses, and their depletion may contribute to mucosal inflammation and barrier dysfunction in IBD.^97,98 Conversely, enrichment of Escherichia–Shigella or Enterococcus may reflect pro-inflammatory activity through lipopolysaccharide-mediated TLR signaling.^99,100 To validate and interpret these associations, further experimental confirmation through in vivo or ex vivo studies is required, as well as the use of next-generation computational frameworks—such as foundation-model-based simulation or causal network modeling—to infer and test host–microbe interaction dynamics.¹⁰¹

Standardization and Reproducibility of Microbiome Assays

A key challenge is the standardization of microbiome assays. Pre-analytical variables such as sample types (feces or biopsies), stool collection methods, preservation techniques, sequencing platforms, and library preparation introduce substantial variability that directly affects downstream feature profiles.

In particular, differences in sequencing chemistry, read length, and bioinformatic preprocessing pipelines can introduce systematic biases in taxonomic resolution and relative abundance estimation.^102,103 Batch effects arising from DNA extraction kits, primer selection, and library preparation protocols have been shown to exceed inter-individual biological variability in some cases, thereby confounding disease-associated microbial signatures across studies.¹⁰⁴

Together, these technical factors interact with sampling- and host-related sources of variability, underscoring the importance of methodological harmonization. Kim et al compared the gut microbiome profiles obtained from the stool, luminal contents, and mucosal biopsies of patients with UC and healthy controls.¹⁰⁵ They found significant differences in both alpha and beta diversities across sample types, with biopsies yielding higher numbers of observed OTUs than stool or lavage samples. Community structures showed a correlation between stool and luminal contents, whereas biopsy samples did not correlate with either. Importantly, UC-control differences were evident only in stool samples, whereas lavage and biopsy samples did not show significant separation, partly due to the limited sample size. Classification analysis achieved AUCs of 0.85 for stool and 0.81 for lavage, underscoring that the sampling strategy critically influences microbiome readouts and diagnostic performance. Kruger et al systematically assessed intra- and inter-individual variability in gut health markers in healthy adults using an optimized fecal sampling and processing workflow.¹⁰⁶ They showed that the variability differed by marker, with some exhibiting substantial within-person variations. Importantly, they demonstrated that optimized pre-processing methods, such as mill homogenization, could significantly reduce technical variability, highlighting that methodological choices can obscure biological signals if not standardized. This underscores the need for assay harmonization to ensure reliable biomarker discovery and provide a robust foundation for ML applications in clinical contexts. Similarly, Clooney et al performed a large longitudinal multicenter study across two continents in patients with IBD and controls, revealing that the geographic site and temporal factors explained a significant share of the microbiome variance, often rivaling or exceeding disease effects.¹⁰⁷ They further observed that temporal stability was reduced in IBD, and that model performance improved when consecutive time points were considered. These findings illustrate how cross-site and temporal heterogeneity can mask disease-associated signatures, and emphasize the importance of standardized methodologies and rigorous external validation in the development of ML-based diagnostics and prognostics for IBD. Unless these sources of heterogeneity are rigorously controlled, models trained in one context are unlikely to be generalized to other clinical settings.¹⁰⁸ This challenge becomes even more pronounced when integrating multi-omics layers, such as metabolomic, transcriptomic, and proteomic data, which adds further complexity to model development and validation.

Model Interpretability and Explainable AI

Another major concern was the interpretability of the model. Algorithms such as RF, boosting algorithms, and neural networks often achieve high predictive performance but can act as “black boxes”.⁹⁶ In clinical practice, the opacity of AI models can undermine the trust of clinicians and complicate regulatory approval. A systematic review by Rosenbacke et al reported that explainable AI can either enhance or diminish trust depending on how explanations are presented.¹⁰⁹ Explanations that are clear, concise, and consistent with clinical reasoning tend to increase clinicians’ confidence in AI systems. Conversely, explanations that are overly technical, complex, inconsistent with clinical logic, or misleading can reduce it. Thus, providing transparent and clinically relevant explanations is essential for successfully integrating AI into healthcare and securing the trust of medical professionals. Rynazal et al applied XAI methods (notably SHAP for local explanations) to colorectal cancer (CRC) classification using gut microbiome data.¹¹⁰ They demonstrated that using local (patient-specific) feature attributes revealed which microbial taxa were most influential for individual predictions and enabled the stratification of patients with CRC into subgroups, thereby improving interpretability and potentially pointing toward mechanistic hypotheses. Yu et al developed a 10-species microbial signature for IBD using an XGBoost-based model (termed XGB-IBD10) and performed external validation.¹¹¹ Importantly, they applied SHAP analysis to explain model predictions, showing how individual taxa influenced the classification decisions. External validation confirmed the robustness of the identified microbial panel, demonstrating that explainability could enhance trust in model outputs and provide mechanistic insights into disease-associated microbes.

Taken together, these studies illustrate that XAI approaches, including feature importance ranking, SHAP values, and attention-based visualization, are critical not only for identifying which microbial taxa, metabolic pathways, or host features drive predictions but also for facilitating clinical adoption and guiding biological hypothesis generation. This process can provide a reliable output for clinical deployment.

Demonstrating Clinical Utility and Path to Adoption

Finally, the demonstration of its clinical utility remains a missing link. Current studies primarily focus on metrics such as the AUC, sensitivity, and specificity. Although these benchmarks are essential for the initial validation, they fail to capture whether such models can improve patient care. To achieve regulatory approval and widespread clinical uptake, predictive frameworks must demonstrate tangible benefits, such as reducing unnecessary colonoscopies, expediting the time to therapy initiation, and lowering healthcare costs. Ultimately, prospective validation and interventional studies are required to bridge the gap between computational promise and real-world clinical impact.

Recent longitudinal research has underscored this need. Al Radi et al investigated the predictive value of gut microbiome signatures for therapy intensification in IBD using a 10-year follow-up cohort.¹¹² Their findings demonstrated that baseline microbial features could predict the need for treatment escalation, highlighting the importance of long-term monitoring and the value of microbiome-informed risk stratification for guiding clinical decision making. This study illustrated how longitudinal datasets provide a rigorous framework for evaluating the durability and clinical relevance of microbiome-derived predictors. In parallel, Massaro et al presented the design of the OPTIMIST study, a prospective, longitudinal, observational pilot study conducted in Canada to investigate gut microbiome predictors of advanced therapy response in CD.¹¹³ The study enrolled patients initiating biological treatment, as well as non-IBD controls, to link baseline microbiome composition and temporal dynamics to therapeutic outcomes. By prospectively monitoring patients over time, the OPTIMIST study established a framework for evaluating microbiome-based predictors in real-world settings and supporting their eventual clinical translation.

In summary, ML and DL hold considerable promise for translating microbiome research into diagnostic and prognostic tools for IBD. However, clinical translation will require overcoming persistent challenges related to reproducibility across cohorts, generalizability to diverse patient populations, interpretability of results, and addressing data-sharing constraints. Progress in explainable AI, development of standardized pipelines, and adherence to rigorous reporting and evaluation guidelines will be crucial in bridging the gap between proof-of-concept studies and real-world patient care.

Conclusion and Future Perspectives

ML and DL have transformed IBD microbiome research from descriptive community profiling to predictive modeling. Alongside DA analyses, these approaches have enabled the classification of disease versus health, the differentiation of CD from UC, and the identification of microbial and metabolic biomarkers with diagnostic and prognostic significance (Figure 1). Studies employing RF, boosting frameworks, and ensemble models have demonstrated that microbiome-informed classifiers can approach or even surpass traditional biomarkers, and explainable AI methods are beginning to provide clearer insights into the microbial and functional drivers of diseases.

Figure 1 Evolution of analytical approaches for characterizing gut microbiota alterations in inflammatory bowel disease (IBD). An overview of analytical strategies for IBD illustrates the progression from traditional statistics to next-generation artificial intelligence (AI) frameworks. Early differential abundance (DA) analyses using 16S rRNA or metagenomic data identified disease-associated taxa; however, these were often limited by univariate contrasts and reproducibility issues. With the increasing data complexity of multi-omics approaches, machine learning (ML) approaches such as random forests, support vector machines (SVMs), and boosting algorithms have enabled robust disease classification, biomarker panel identification, and performance evaluation through receiver operating characteristic (ROC) curves, cross-validation, and feature importance scores. Recently, deep learning (DL) frameworks have been applied to capture nonlinear relationships in high-dimensional multi-omics datasets, providing latent feature embeddings and integrative predictive models. Looking ahead, foundation models trained on large-scale host and microbial genomic resources (eg, Microbial General Model for microbiome and AlphaGenome for host genetics) are emerging as transferable backbones that can be fine-tuned for diverse downstream tasks, from diagnosis and subtype prediction to therapy response modeling and prognosis.

Despite this progress, significant barriers remain before clinical adoption can be achieved. Variability in stool collection, sequencing platforms, and preprocessing pipelines continues to hinder reproducibility, and external validation often reveals a decline in model performance when applied across independent cohorts. Dependence on complex “black-box” algorithms also makes it more difficult for clinicians to trust and regulators to approve. Furthermore, most studies are retrospective, with few prospective or interventional studies directly evaluating whether microbiome-based ML models can improve patient outcomes. These issues emphasize the need to connect computational advances with the development of practical tools for precision medicine.

Looking ahead, progress will be fastest if efforts focus on: (1) rigorous validation, prioritizing pre-registered pipelines, locked models, and multi-site external testing; (2) strong benchmarking, with standardized preprocessing, shared baselines, and challenge datasets covering stool and mucosal samples, platforms, and regions; (3) biologically based interpretability, pairing explanation methods with wet-lab or orthogonal assays to verify mechanistic plausibility; (4) longitudinal and multimodal integration, combining metagenomics, metabolomics, proteomics, and host features to model disease trajectories rather than single time points; and (5) next-generation AI, where emerging genome foundation models—both microbiome-focused and host-centric—are developed under strict evaluation and compared directly to robust classical baselines. These steps will help ML move from promising prototypes to dependable clinical tools that guide diagnosis, monitoring, and treatment decisions for IBD.

Data Sharing Statement

Data sharing not applicable – no new data generated.

Acknowledgments

The figure was generated using BioRender by June-Young Lee (2025) https://BioRender.com/je7gx57. We would like to thank Editage for the English language editing.

Author Contributions

• June-Young Lee (J.Y. Lee): Conceptualization; Methodology; Formal analysis; Software; Data curation; Investigation; Resources; Visualization; Funding acquisition; Writing – original draft; Writing – review & editing.
• Dong Hyun Kim (D.H. Kim): Formal analysis; Software; Validation; Data curation; Visualization; Writing – review & editing.
• Jee-Won Choi (J.W. Choi): Data curation; Investigation; Visualization; Writing – review & editing.
• Minho Shong (M. Shong): Resources; Investigation; Writing – review & editing.
• Chang Kyun Lee (C.K. Lee): Conceptualization; Methodology; Supervision; Project administration; Funding acquisition; Writing – review & editing.

All authors contributed substantially to the work, critically revised the manuscript, and agree to be accountable for all aspects of the work. All authors gave final approval of the version to be published. June-Young Lee and Dong Hyun Kim contributed equally to this work and share first authorship.

Funding

This research was supported by a grant from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute, funded by the Ministry of Health & Welfare, Republic of Korea (grant number: RS-2023-KH135855) and the InnoCORE program of the Ministry of Science and ICT, Republic of Korea (grant number: N10250153).

Disclosure

The authors declare that they have no competing interests.

References

1. Ni J, Wu GD, Albenberg L, Tomov VT. Gut microbiota and IBD: causation or correlation? Nat Rev Gastroenterol Hepatol. 2017;14(10):573–584. doi:10.1038/nrgastro.2017.88

2. Kostic AD, Xavier RJ, Gevers D. The microbiome in inflammatory bowel disease: current status and the future ahead. Gastroenterology. 2014;146(6):1489–1499. doi:10.1053/j.gastro.2014.02.009

3. Shan Y, Lee M, Chang EB. The gut microbiome and inflammatory bowel diseases. Annual Review of Medicine. 2022;73:455–468. doi:10.1146/annurev-med-042320-021020

4. Lloyd-Price J, Arze C, Ananthakrishnan AN, et al. Multi-omics of the gut microbial ecosystem in inflammatory bowel diseases. Nature. 2019;569(7758):655–662. doi:10.1038/s41586-019-1237-9

5. Morgan XC, Tickle TL, Sokol H, et al. Dysfunction of the intestinal microbiome in inflammatory bowel disease and treatment. Genome Biol. 2012;13(9):R79. doi:10.1186/gb-2012-13-9-r79

6. Ning L, Zhou Y-L, Sun H, et al. Microbiome and metabolome features in inflammatory bowel disease via multi-omics integration analyses across cohorts. Nat Commun. 2023;14(1):7135. doi:10.1038/s41467-023-42788-0

7. Zhang Y, Thomas JP, Korcsmaros T, Gul L. Integrating multi-omics to unravel host-microbiome interactions in inflammatory bowel disease. Cell Rep Med. 2024;5(9):101738. doi:10.1016/j.xcrm.2024.101738

8. Hu S, Bourgonje AR, Gacesa R, et al. Mucosal host-microbe interactions associate with clinical phenotypes in inflammatory bowel disease. Nat Commun. 2024;15(1):1470. doi:10.1038/s41467-024-45855-2

9. Jayasooriya N, Baillie S, Blackwell J, et al. Systematic review with meta-analysis: time to diagnosis and the impact of delayed diagnosis on clinical outcomes in inflammatory bowel disease. Aliment. Pharmacol. Ther. 2023;57(6):635–652. doi:10.1111/apt.17370

10. Zhang SY, Lin Y. Addressing diagnostic delays in inflammatory bowel diseases in Germany. World J Gastroenterol. 2024;30(41):4484–4489. doi:10.3748/wjg.v30.i41.4484

11. Henriksen M, Jahnsen J, Lygren I, et al. C-reactive protein: a predictive factor and marker of inflammation in inflammatory bowel disease. Results from a prospective population-based study. Gut. 2008;57(11):1518–1523. doi:10.1136/gut.2007.146357

12. Walsham NE, Sherwood RA. Fecal calprotectin in inflammatory bowel disease. Clin Exp Gastroenterol. 2016;9:21–29. doi:10.2147/ceg.S51902

13. Clough J, Colwill M, Poullis A, Pollok R, Patel K, Honap S. Biomarkers in inflammatory bowel disease: a practical guide. Ther Adv Gastroenterol. 2024;17:17562848241251600. doi:10.1177/17562848241251600

14. Verstockt B, Noor NM, Marigorta UM, Pavlidis P, Deepak P, Ungaro RC. Results of the seventh scientific workshop of ECCO: precision medicine in IBD-disease outcome and response to therapy. J Crohns Colitis. 2021;15(9):1431–1442. doi:10.1093/ecco-jcc/jjab050

15. Chen L, Zhang C, Niu R, et al. Multi-omics biomarkers for predicting efficacy of biologic and small-molecule therapies in adults with inflammatory bowel disease: a systematic review. United Eur Gastroenterol J. 2025;13(4):517–530. doi:10.1002/ueg2.12720

16. Mallick H, Rahnavard A, McIver LJ, et al. Multivariable association discovery in population-scale meta-omics studies. PLoS Comput Biol. 2021;17(11):e1009442. doi:10.1371/journal.pcbi.1009442

17. Segata N, Izard J, Waldron L, et al. Metagenomic biomarker discovery and explanation. Genome Biol. 2011;12(6):R60. doi:10.1186/gb-2011-12-6-r60

18. Hernández Medina R, Kutuzova S, Nielsen KN, et al. Machine learning and deep learning applications in microbiome research. ISME Commun. 2022;2(1):98. doi:10.1038/s43705-022-00182-9

19. Nearing JT, Douglas GM, Hayes MG, et al. Microbiome differential abundance methods produce different results across 38 datasets. Nat Commun. 2022;13(1):342. doi:10.1038/s41467-022-28034-z

20. Halfvarson J, Brislawn CJ, Lamendella R, et al. Dynamics of the human gut microbiome in inflammatory bowel disease. Nat Microbiol. 2017;2(5):17004. doi:10.1038/nmicrobiol.2017.4

21. Franzosa EA, Sirota-Madi A, Avila-Pacheco J, et al. Gut microbiome structure and metabolic activity in inflammatory bowel disease. Nat Microbiol. 2019;4(2):293–305. doi:10.1038/s41564-018-0306-4

22. Tarca AL, Carey VJ, X-w C, Romero R, Drăghici S. Machine Learning and Its Applications to Biology. PLoS Comput Biol. 2007;3(6):e116. doi:10.1371/journal.pcbi.0030116

23. Marcos-Zambrano LJ, Karaduzovic-Hadziabdic K, Loncar Turukalo T, et al. Applications of machine learning in human microbiome studies: a review on feature selection, biomarker identification, disease prediction and treatment. Review. Front Microbiol. 2021;Volume 12. doi:10.3389/fmicb.2021.634511

24. Zheng J, Sun Q, Zhang M, et al. Noninvasive, microbiome-based diagnosis of inflammatory bowel disease. Nat Med. 2024;30(12):3555–3567. doi:10.1038/s41591-024-03280-4

25. Khachatryan L, Xiang Y, Ivanov A, et al. Results and lessons learned from the sbv IMPROVER metagenomics diagnostics for inflammatory bowel disease challenge. Sci Rep. 2023;13(1):6303. doi:10.1038/s41598-023-33050-0

26. Caminero A, Tropini C, Valles-Colomer M, et al. Credible inferences in microbiome research: ensuring rigour, reproducibility and relevance in the era of AI. Nat Rev Gastroenterol Hepatol. 2025;22(11):788–803. doi:10.1038/s41575-025-01100-9

27. Gilbert JA, Azad MB, Bäckhed F, et al. Clinical translation of microbiome research. Nat Med. 2025;31(4):1099–1113. doi:10.1038/s41591-025-03615-9

28. Ma X, Lu X, Zhang W, et al. Gut microbiota in the early stage of Crohn’s disease has unique characteristics. Gut Pathog. 2022;14(1):46. doi:10.1186/s13099-022-00521-0

29. Nickols WA, Kuntz T, Shen J, et al. MaAsLin 3: refining and extending generalized multivariable linear models for meta-omic association discovery. bioRxiv. 2024:2024.12.13.628459. doi:10.1101/2024.12.13.628459

30. Lin H, Peddada SD. Analysis of compositions of microbiomes with bias correction. Nat Commun. 2020;11(1):3514. doi:10.1038/s41467-020-17041-7

31. Lin H, Peddada SD. Multigroup analysis of compositions of microbiomes with covariate adjustments and repeated measures. Nat Method. 2024;21(1):83–91. doi:10.1038/s41592-023-02092-7

32. Chen W, Li Y, Wang W, et al. Enhanced microbiota profiling in patients with quiescent Crohn’s disease through comparison with paired healthy first-degree relatives. Cell Rep Med. 2024;5(7):101624. doi:10.1016/j.xcrm.2024.101624

33. Rosso AD, Aguilera P, Quesada S, et al. Comprehensive phenotyping in inflammatory bowel disease: search for biomarker algorithms in the transkingdom interactions context. Microorganisms. 2022;10(11):2190. doi:10.3390/microorganisms10112190

34. Gloor GB. ALDEx 2: ANOVA-Like Differential Expression tool for compositional data. 2016.

35. Zhou H, He K, Chen J, Zhang X. LinDA: linear models for differential abundance analysis of microbiome compositional data. Genome Biol. 2022;23(1):95. doi:10.1186/s13059-022-02655-5

36. Vich Vila A, Imhann F, Collij V, et al. Gut microbiota composition and functional changes in inflammatory bowel disease and irritable bowel syndrome. Sci Transl Med. 2018;10(472). doi:10.1126/scitranslmed.aap8914

37. Akiyama S, Nishijima S, Kojima Y, et al. Multi-biome analysis identifies distinct gut microbial signatures and their crosstalk in ulcerative colitis and Crohn’s disease. Nat Commun. 2024;15(1):10291. doi:10.1038/s41467-024-54797-8

38. Tun HM, Peng Y, Massimino L, et al. Gut virome in inflammatory bowel disease and beyond. Gut. 2024;73(2):350–360. doi:10.1136/gutjnl-2023-330001

39. Norman Jason M, Handley Scott A, Baldridge Megan T, et al. Disease-specific alterations in the enteric virome in inflammatory bowel disease. Cell. 2015;160(3):447–460. doi:10.1016/j.cell.2015.01.002

40. Hsu C, Ghannoum M, Cominelli F, Martino LD. Mycobiome and inflammatory bowel disease: role in disease pathogenesis, current approaches and novel nutritional-based therapies. Inflamm Bowel Dis. 2023;29(3):470–479. doi:10.1093/ibd/izac156

41. Zhang X, Deeke SA, Ning Z, et al. Metaproteomics reveals associations between microbiome and intestinal extracellular vesicle proteins in pediatric inflammatory bowel disease. Nat Commun. 2018;9(1):2873. doi:10.1038/s41467-018-05357-4

42. Li M, Yang L, Mu C, et al. Gut microbial metabolome in inflammatory bowel disease: from association to therapeutic perspectives. Comput Struct Biotechnol J. 2022;20:2402–2414. doi:10.1016/j.csbj.2022.03.038

43. Liu L, Wang G, Safo SE. Multivariate sparse functional linear discriminant analysis: an application to inflammatory bowel disease classification. arXiv preprint, arXiv. 2025.

44. Gloor GB, Macklaim JM, Pawlowsky-Glahn V, Egozcue JJ. Microbiome datasets are compositional: and this is not optional. Mini review. Front Microbiol. 2017;8. doi:10.3389/fmicb.2017.02224

45. Silverman JD, Washburne AD, Mukherjee S, David LA. A phylogenetic transform enhances analysis of compositional microbiota data. eLife. 2017;6:e21887. doi:10.7554/eLife.21887

46. Hodgkiss R, Acharjee A. Unravelling metabolite-microbiome interactions in inflammatory bowel disease through AI and interaction-based modelling. Biochimica et Biophysica Acta. 2025;1871(3):167618. doi:10.1016/j.bbadis.2024.167618

47. Manandhar I, Alimadadi A, Aryal S, Munroe PB, Joe B, Cheng X. Gut microbiome-based supervised machine learning for clinical diagnosis of inflammatory bowel diseases. Am J Physiol Gastrointest Liver Physiol. 2021;320(3):G328–g337. doi:10.1152/ajpgi.00360.2020

48. Papoutsoglou G, Tarazona S, Lopes MB, et al. Machine learning approaches in microbiome research: challenges and best practices. Review. Front Microbiol. 2023;14. doi:10.3389/fmicb.2023.1261889

49. Lee S, Lee I. Comprehensive assessment of machine learning methods for diagnosing gastrointestinal diseases through whole metagenome sequencing data. Gut Microbes. 2024;16(1):2375679. doi:10.1080/19490976.2024.2375679

50. Liñares-Blanco J, Fernandez-Lozano C, Seoane JA, López-Campos G. Machine learning based microbiome signature to predict inflammatory bowel disease subtypes. original research. Front Microbiol. 2022;13. doi:10.3389/fmicb.2022.872671

51. Kim H, Na JE, Kim S, et al. A machine learning-based diagnostic model for crohn’s disease and ulcerative colitis utilizing fecal microbiome analysis. Microorganisms. 2023;12(1):36. doi:10.3390/microorganisms12010036

52. Ha SM, Lee K, Kim GH, Hurych J, Cinek O, Shim JO. Gut-microbiota-based ensemble model predicts prognosis of pediatric inflammatory bowel disease. Iscience. 2024;27(12):111442. doi:10.1016/j.isci.2024.111442

53. Jacobs JP, Sauk JS, Ahdoot AI, et al. Microbial and metabolite signatures of stress reactivity in ulcerative colitis patients in clinical remission predict clinical flare risk. Inflamm Bowel Dis. 2023;30(3):336–346. doi:10.1093/ibd/izad185

54. Kim HS, Oh SJ, Kim BK, et al. Dysbiotic signatures and diagnostic potential of gut microbial markers for inflammatory bowel disease in Korean population. Sci Rep. 2024;14(1):23701. doi:10.1038/s41598-024-74002-6

55. Lee Y, Cappellato M, Di Camillo B. Machine learning-based feature selection to search stable microbial biomarkers: application to inflammatory bowel disease. Gigascience. 2022;12. doi:10.1093/gigascience/giad083

56. Abbas M, Matta J, Le T, et al. Biomarker discovery in inflammatory bowel diseases using network-based feature selection. PLoS One. 2019;14(11):e0225382. doi:10.1371/journal.pone.0225382

57. Breiman L. Random Forests. Mach Learn. 2001;45(1):5–32. doi:10.1023/A:1010933404324

58. Statnikov A, Henaff M, Narendra V, et al. A comprehensive evaluation of multicategory classification methods for microbiomic data. Microbiome. 2013;1(1):11. doi:10.1186/2049-2618-1-11

59. Knights D, Costello EK, Knight R. Supervised classification of human microbiota. FEMS Microbiol Rev. 2011;35(2):343–359. doi:10.1111/j.1574-6976.2010.00251.x

60. Li L, Deng X, Wang S, Huang T. Integrating traditional omics and machine learning approaches to identify microbial biomarkers and therapeutic targets in pediatric inflammatory bowel disease. Front Pharmacol. 2025;16:1545392. doi:10.3389/fphar.2025.1545392

61. Li L, Liu Z, Qin J, et al. Constructing inflammatory bowel disease diagnostic models based on k-mer and machine learning. Front Microbiol. 2025;16:1578005. doi:10.3389/fmicb.2025.1578005

62. Gonzalez CG, Mills RH, Zhu Q, et al. Location-specific signatures of Crohn’s disease at a multi-omics scale. Microbiome. 2022;10(1):133. doi:10.1186/s40168-022-01331-x

63. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20(3):273–297. doi:10.1007/BF00994018

64. Jiang Y, Luo J, Huang D, Liu Y, Li DD. Machine learning advances in microbiology: a review of methods and applications. Front Microbiol. 2022;13:925454. doi:10.3389/fmicb.2022.925454

65. Jerome HF. Greedy function approximation: a gradient boosting machine. Ann Stat. 2001;29(5):1189–1232. doi:10.1214/aos/1013203451

66. Przymus P, Rykaczewski K, Martín-Segura A, et al. Deep learning in microbiome analysis: a comprehensive review of neural network models. Rev Front Microbiol. 2025;15. doi:10.3389/fmicb.2024.1516667

67. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436–444. doi:10.1038/nature14539

68. Polikar R. Ensemble based systems in decision making. IEEE Circuits Syst Mag. 2006;6(3):21–45. doi:10.1109/MCAS.2006.1688199

69. Dudek NK, Chakhvadze M, Kobakhidze S, Kantidze O, Gankin Y. Supervised machine learning for microbiomics: bridging the gap between current and best practices. Mach Learn Appl. 2024;18:100607. doi:10.1016/j.mlwa.2024.100607

70. Topçuoğlu BD, Lesniak NA, MTt R, Wiens J, Schloss PD. A framework for effective application of machine learning to microbiome-based classification problems. mBio. 2020;11(3). doi:10.1128/mBio.00434-20

71. Kubinski R, Djamen-Kepaou JY, Zhanabaev T, et al. Benchmark of data processing methods and machine learning models for gut microbiome-based diagnosis of inflammatory bowel disease. Front Genet. 2022;13:784397. doi:10.3389/fgene.2022.784397

72. Lin H, Peddada SD. Analysis of microbial compositions: a review of normalization and differential abundance analysis. Npj Biofilms Microbi. 2020;6(1):60. doi:10.1038/s41522-020-00160-w

73. Pan AY. Statistical analysis of microbiome data: the challenge of sparsity. Curr. Opin. Endocr. Metab. Res. 2021;19:35–40. doi:10.1016/j.coemr.2021.05.005

74. Zand A, Stokes Z, Sharma A, van Deen WK, Hommes D. Artificial intelligence for Inflammatory Bowel Diseases (IBD); accurately predicting adverse outcomes using machine learning. Dig Dis Sci. 2022;67(10):4874–4885. doi:10.1007/s10620-022-07506-8

75. Oh M, Zhang L. DeepMicro: deep representation learning for disease prediction based on microbiome data. Sci Rep. 2020;10(1):6026. doi:10.1038/s41598-020-63159-5

76. Jain S, Safo SE. DeepIDA-GRU: a deep learning pipeline for integrative discriminant analysis of cross-sectional and longitudinal multiview data with applications to inflammatory bowel disease classification. Briefings Bioinf 2024;25(4). doi:10.1093/bib/bbae339

77. Reiman D, Metwally A, Yang D. Using convolutional neural networks to explore the microbiome. Annu Int Conf IEEE Eng Med Biol Soc. 2017;2017:4269–4272. doi:10.1109/embc.2017.8037799

78. Grazioli F, Siarheyeu R, Alqassem I, Henschel A, Pileggi G, Meiser A. Microbiome-based disease prediction with multimodal variational information bottlenecks. PLoS Comput Biol. 2022;18(4):e1010050. doi:10.1371/journal.pcbi.1010050

79. Zhang H, Zhang Y, Kang Z, Song L, Yang R, Ning K. MGM as a large-scale pretrained foundation model for microbiome analyses in diverse contexts. bioRxiv. 2025. doi:10.1101/2024.12.30.630825

80. Avsec Ž, Latysheva N, Cheng J, et al. AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model. bioRxiv. 2025. doi:10.1101/2025.06.25.661532

81. Dalla-Torre H, Gonzalez L, Mendoza-Revilla J, et al. Nucleotide transformer: building and evaluating robust foundation models for human genomics. Nat Method. 2025;22(2):287–297. doi:10.1038/s41592-024-02523-z

82. Sazal M, Stebliankin V, Mathee K, Yoo C, Narasimhan G. Causal effects in microbiomes using interventional calculus. Sci Rep. 2021;11(1):5724. doi:10.1038/s41598-021-84905-3

83. Lyu R, Qu Y, Divaris K, Wu D. Methodological considerations in longitudinal analyses of microbiome data: a comprehensive review. Genes. 2024;15(1):51. doi:10.3390/genes15010051

84. Ruiz-Perez D, Lugo-Martinez J, Bourguignon N, et al. Dynamic bayesian networks for integrating multi-omics time series microbiome data. Msystems. 2021;6(2). doi:10.1128/mSystems.01105-20

85. Argelaguet R, Velten B, Arnol D, et al. Multi-omics factor analysis-a framework for unsupervised integration of multi-omics data sets. Mol Syst Biol. 2018;14(6):e8124. doi:10.15252/msb.20178124

86. Singh A, Shannon CP, Gautier B, et al. DIABLO: an integrative approach for identifying key molecular drivers from multi-omics assays. Bioinformatics. 2019;35(17):3055–3062. doi:10.1093/bioinformatics/bty1054

87. Muller E, Shiryan I, Borenstein E. Multi-omic integration of microbiome data for identifying disease-associated modules. Nat Commun. 2024;15(1):2621. doi:10.1038/s41467-024-46888-3

88. Sharma C, Sharma S, Sharma K, Sethi GK, Chen H-Y. Exploring explainable AI: a bibliometric analysis. Discov Appl Sci. 2024;6(11):615. doi:10.1007/s42452-024-06324-z

89. Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. presented at: Proceedings of the 31st International Conference on Neural Information Processing Systems; 2017; Long Beach, California, USA.

90. Novielli P, Romano D, Magarelli M, et al. Explainable artificial intelligence for microbiome data analysis in colorectal cancer biomarker identification. Original Research. Front Microbiol. 2024;15. doi:10.3389/fmicb.2024.1348974

91. Onwuka S, Bravo-Merodio L, Gkoutos GV, Acharjee A. Explainable AI-prioritized plasma and fecal metabolites in inflammatory bowel disease and their dietary associations. Iscience. 2024;27(7):110298. doi:10.1016/j.isci.2024.110298

92. Kim M, Gim D, Kim S, et al. From dysbiosis to prediction: AI-powered microbiome insights into IBD and CRC. Gastroenterol Insights. 2025;16(3):34. doi:10.3390/gastroent16030034

93. Stafford IS, Gosink MM, Mossotto E, Ennis S, Hauben M. A systematic review of artificial intelligence and machine learning applications to inflammatory bowel disease, with practical guidelines for interpretation. Inflamm Bowel Dis. 2022;28(10):1573–1583. doi:10.1093/ibd/izac115

94. Tariq R, Afzali A. Artificial intelligence in inflammatory bowel disease: innovations in diagnosis, monitoring, and personalized care. Therap Adv Gastroenterol. 2025;18:17562848251357407. doi:10.1177/17562848251357407

95. De Deo D, Dal Buono A, Gabbiadini R, et al. Digital biomarkers and artificial intelligence: a new frontier in personalized management of inflammatory bowel disease. Rev Front Immunol. 2025;16. doi:10.3389/fimmu.2025.1637159

96. Tourab H, López Pérez L, Arroyo-Gallego P, et al. The use of machine learning and explainable artificial intelligence in gut microbiome research: a scoping review. 2024.

97. Lenoir M, Martín R, Torres-Maravilla E, et al. Butyrate mediates anti-inflammatory effects of Faecalibacterium prausnitzii in intestinal epithelial cells through Dact3. Gut Microbes. 2020;12(1):1–16. doi:10.1080/19490976.2020.1826748

98. Zhang Z, Zhang H, Chen T, Shi L, Wang D, Tang D. Regulatory role of short-chain fatty acids in inflammatory bowel disease. Cell Commun Signal. 2022;20(1):64. doi:10.1186/s12964-022-00869-5

99. Mirsepasi-Lauridsen HC, Vallance BA, Krogfelt KA, Petersen AM. Escherichia coli pathobionts associated with inflammatory bowel disease. Clin Microbiol Rev. 2019;32(2). doi:10.1128/cmr.00060-18

100. Giambra V, Pagliari D, Rio P, et al. Gut microbiota, inflammatory bowel disease, and cancer: the role of guardians of innate immunity. Cells. 2023;12(22):2654. doi:10.3390/cells12222654

101. Jansma J, El Aidy S. Understanding the host-microbe interactions using metabolic modeling. Microbiome. 2021;9(1):16. doi:10.1186/s40168-020-00955-1

102. Nearing JT, Comeau AM, Langille MGI. Identifying biases and their potential solutions in human microbiome studies. Microbiome. 2021;9(1):113. doi:10.1186/s40168-021-01059-0

103. Allali I, Arnold JW, Roach J, et al. A comparison of sequencing platforms and bioinformatics pipelines for compositional analysis of the gut microbiome. BMC Microbiol. 2017;17(1):194. doi:10.1186/s12866-017-1101-8

104. Xu Z, Yeoh YK, Tun HM, et al. Variation in the metagenomic analysis of fecal microbiome composition calls for a standardized operating approach. Microbiol Spectr. 2024;12(12):e0151624. doi:10.1128/spectrum.01516-24

105. Kim D, Jung J-Y, Oh H-S, et al. Comparison of sampling methods in assessing the microbiome from patients with ulcerative colitis. BMC Gastroenterol. 2021;21(1):396. doi:10.1186/s12876-021-01975-3

106. Kruger K, Myeonghyun Y, van der Wielen N, et al. Evaluation of inter- and intra-variability in gut health markers in healthy adults using an optimised faecal sampling and processing method. Sci Rep. 2024;14(1):24580. doi:10.1038/s41598-024-75477-z

107. Clooney AG, Eckenberger J, Laserna-Mendieta E, et al. Ranking microbiome variance in inflammatory bowel disease: a large longitudinal intercontinental study. Gut. 2021;70(3):499–510. doi:10.1136/gutjnl-2020-321106

108. Huang J, Zhu X, Ma Y, et al. Machine learning in the differential diagnosis of ulcerative colitis and Crohn’s disease: a systematic review. Translat Gastroenterol Hepatol. 2025;10:10. doi:10.21037/tgh-24-87

109. Rosenbacke R, Å M, McKee M, Stuckler D. How explainable artificial intelligence can increase or decrease clinicians’ trust in AI applications in health care: systematic review. JMIR AI. 2024;3:e53207. doi:10.2196/53207

110. Rynazal R, Fujisawa K, Shiroma H, et al. Leveraging explainable AI for gut microbiome-based colorectal cancer classification. Genome Biol. 2023;24(1):21. doi:10.1186/s13059-023-02858-4

111. Yu S, Li J, Ye Z, et al. Identification of a 10-species microbial signature of inflammatory bowel disease by machine learning and external validation. Cell Regen. 2025;14(1):32. doi:10.1186/s13619-025-00246-w

112. Al Radi ZMA, Prins FM, Collij V, et al. Exploring the predictive value of gut microbiome signatures for therapy intensification in patients with inflammatory bowel disease: a 10-year follow-up study. Inflamm Bowel Dis. 2024;30(10):1642–1653. doi:10.1093/ibd/izae064

113. Massaro CA, Meade S, Lemarié FL, et al. Gut microbiome predictors of advanced therapy response in Crohn’s disease: protocol for the OPTIMIST prospective, longitudinal, observational pilot study in Canada. BMJ Open. 2025;15(3):e094280. doi:10.1136/bmjopen-2024-094280