Issa IA, Issa T. Role of artificial intelligence in irritable bowel syndrome: A narrative review. Artif Intell Gastroenterol 2026; 7(3): 127488 [DOI: 10.35712/aig.127488]
Corresponding Author of This Article
Iyad A Issa, MD, Faculty of Health Sciences, Lebanese University, Hadath, Beirut 0000, Beyrouth, Lebanon. iyadissa71@gmail.com
Research Domain of This Article
Gastroenterology & Hepatology
Article-Type of This Article
review-article
Open-Access Policy of This Article
This article is an open-access article which was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/
Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA
Share the Article
Issa IA, Issa T. Role of artificial intelligence in irritable bowel syndrome: A narrative review. Artif Intell Gastroenterol 2026; 7(3): 127488 [DOI: 10.35712/aig.127488]
Author contributions: Issa I conceptualized and designed the review, performed the literature search, and drafted and critically revised the manuscript; Issa T contributed to the literature search, data interpretation, and drafting of the manuscript; and both authors have read and approved the final version of the manuscript.
AI contribution statement: During the preparation of this work, the authors used Claude AI for English-language polishing. The authors reviewed and edited the output as needed and take full responsibility for the content of the article.
Conflict-of-interest statement: The authors declare that they have no conflicts of interest.
Corresponding author: Iyad A Issa, MD, Faculty of Health Sciences, Lebanese University, Hadath, Beirut 0000, Beyrouth, Lebanon. iyadissa71@gmail.com
Received: September 2, 2026 Revised: September 18, 2026 Accepted: October 28, 2026 Published online: November 8, 2026 Processing time: 65 Days and 11.3 Hours
Abstract
Irritable bowel syndrome (IBS) is a prevalent yet incompletely understood functional gastrointestinal disorder whose estimated prevalence depends heavily on the diagnostic criteria applied, falling from approximately 9% under Rome III to approximately 4% under Rome IV. Characterized by recurrent abdominal pain and altered bowel habits, it substantially impairs quality of life and imposes a considerable economic burden. Its pathophysiology is complex and multifactorial, involving visceral hypersensitivity, altered gastrointestinal motility, immune activation, gut-brain axis dysregulation, and intestinal dysbiosis. This narrative review examines the potential of artificial intelligence (AI) to refine the diagnosis, subclassification, and management of IBS and related functional gastrointestinal disorders. We discuss the principal AI methodologies - machine learning, deep learning, and natural language processing - and how they enable the integration of diverse data sources, including patient-reported outcomes, microbiome profiles, and multi-omics data. Early studies suggest AI can support the differentiation of IBS from organic disease, the identification of biologically distinct microbiota subgroups, and the prediction of response to the low-fermentable oligosaccharides, disaccharides, monosaccharides and polyols diet. These findings, however, derive largely from small, single-center cohorts with heterogeneous methods and without external validation; a 2025 systematic review identified only six eligible studies comprising 1366 participants and found the evidence too heterogeneous to permit meta-analysis. We further identify domains in which AI is frequently invoked but not yet evidenced in IBS, including machine-learning analysis of longitudinal patient-collected symptom data and of psychological comorbidity. Substantial challenges remain around data standardization, external and temporal validation, model interpretability, and ethical and regulatory considerations. Progress will depend less on the development of further models than on standardized data collection, multi-center validation, and reporting against AI-specific standards such as transparent reporting of a multivariable prediction model for individual prognosis or diagnosis + AI.
Core Tip: Enthusiasm for artificial intelligence in irritable bowel syndrome has outpaced its evidence base. Across diagnosis, microbiota subtyping and treatment prediction, the supporting studies are small, single-center and largely externally unvalidated; a 2025 systematic review found only six eligible studies and could not pool them. Several widely repeated capabilities - machine-learning analysis of longitudinal symptom data, and of psychological comorbidity - have no peer-reviewed IBS evidence that we were able to identify. The field’s priority should therefore shift from developing further models toward standardized data collection, multi-center external validation, and reporting against artificial intelligence-specific standards.
Citation: Issa IA, Issa T. Role of artificial intelligence in irritable bowel syndrome: A narrative review. Artif Intell Gastroenterol 2026; 7(3): 127488
Irritable bowel syndrome (IBS) exemplifies the diagnostic and mechanistic complexity of the functional gastrointestinal disorders (FGIDs), now more commonly termed disorders of gut-brain interaction (DGBI); the two terms are used interchangeably here, consistent with current Rome Foundation nomenclature. It is characterized by recurrent abdominal pain and altered bowel habits whose symptoms cannot be attributed to another medical condition after appropriate evaluation[1]. Reported prevalence depends substantially on the diagnostic criteria applied: Pooled global estimates are approximately 9% under Rome III but fall to approximately 4% under the more restrictive Rome IV criteria[2], and the Rome Foundation Global Study likewise reported a Rome IV prevalence of 4.1% compared with 10.1% under Rome III within the same dataset[3]. This contraction is itself instructive for a review concerned with diagnostic refinement, since it reflects the sensitivity of case definition to criteria rather than any change in underlying disease burden. The economic burden nonetheless remains substantial, with reported annual direct costs of United States $1562-7547 and indirect costs of United States $791-7737 per patient[4].
The pathophysiology of IBS remains incompletely understood, involving multifactorial mechanisms including visceral hypersensitivity, altered gastrointestinal motility, immune activation, gut-brain axis dysregulation, and intestinal dysbiosis[5]. This etiological heterogeneity translates into diverse clinical phenotypes, conventionally categorized as IBS with predominant constipation, predominant diarrhea, mixed bowel habits (IBS-M), and unclassified IBS[1]. The symptom-based diagnostic approach, while clinically practical, often fails to capture the underlying biological heterogeneity, contributing to suboptimal treatment responses[6].
Artificial intelligence (AI), including machine learning (ML), deep learning (DL), and natural language processing (NLP), has advanced multiple medical disciplines by enabling the analysis of complex, high-dimensional datasets that exceed human cognitive capacity[7]. In gastroenterology, AI applications have demonstrated notable efficacy in endoscopic image recognition, including polyp detection and optical prediction of histology[8,9]. More recently, AI methodologies have begun to be applied to functional disorders, where the integration of patient-reported outcomes, microbiome data, metabolomics, and clinical parameters offers new opportunities for precision medicine[10].
LITERATURE SELECTION STRATEGY
This is a narrative review, reported with reference to the SANRA quality criteria for narrative review articles[11]. We searched PubMed/MEDLINE and EMBASE from January 2000 to March 31, 2026, combining the terms (“irritable bowel syndrome” OR “functional gastrointestinal disorder” OR “disorders of gut-brain interaction”) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “natural language processing” OR “microbiome”), and hand-searched the reference lists of included articles. Search strings were adapted to each database, using Medical Subject Headings together with title and abstract terms in PubMed/MEDLINE and the corresponding Emtree terms with equivalent free-text terms in EMBASE. Titles and abstracts were screened by the first author, full texts of potentially eligible records were assessed by both authors, and disagreements were resolved by discussion. Eligible sources were English-language primary studies, systematic reviews, and methodological or reporting-guideline references addressing AI methods in IBS or related FGIDs. Conference abstracts were excluded, as were preprints except where a methodological source had no peer-reviewed equivalent. No minimum sample size was applied, but cohort size is reported throughout because most available studies are small. Because this is a narrative rather than a systematic review, articles were selected for relevance and methodological illustrativeness rather than by formal protocol, and neither risk-of-bias scoring nor quantitative synthesis was undertaken; we note this as a limitation. Throughout, we distinguish three grades of evidence: Direct IBS evidence, from studies conducted in IBS populations; indirect evidence, extrapolated from other conditions or from generic methodological work; and proposed applications, which are conceptual and have not been tested in IBS. Claims are labelled accordingly in the text, and the evidence grade of each study is recorded in Table 1. The review examines diagnostic refinement, subtype classification, treatment prediction, and microbiome-based interventions, and critically appraises the methodological approaches, clinical applications, and translational challenges that currently define this emerging field. Figure 1 summarizes the overall data-to-decision framework discussed in this review.
Figure 1 Data-to-decision framework for the application of artificial intelligence in irritable bowel syndrome and related functional gastrointestinal disorders.
Diverse data sources - patient-reported outcomes, clinical and laboratory features, microbiome profiles, and multi-omics data - are integrated and analyzed with machine-learning, deep-learning, and natural language processing methods to support diagnostic refinement, biological subtyping, and personalized treatment selection. This figure is a conceptual framework summarizing potential applications; none of the depicted capabilities is validated for routine clinical use in irritable bowel syndrome. AI: Artificial intelligence; IBS: Irritable bowel syndrome; EHR: Electronic health record.
Table 1 Summary of representative artificial intelligence studies in irritable bowel syndrome, with evidence grade.
AI encompasses several families of methods that differ in the data they require and in the questions they can answer. Table 2 summarizes the principal approaches, their characteristic applications, and the extent to which each has actually been deployed in IBS. These methods are described in detail elsewhere[7,12]; here we emphasize the distinction that matters most clinically. Supervised learning requires labelled outcomes and therefore inherits whatever uncertainty exists in the label[13], whereas unsupervised methods impose no labels and may generate subgroups that correspond to no recognized clinical category[14].
Table 2 Artificial intelligence methods and their current status in irritable bowel syndrome.
Method family
Core principle
Typical application
Current status in IBS
Supervised learning (random forest, support vector machine, gradient boosting)
Learns a mapping from labelled examples to a defined outcome
This distinction is consequential in IBS, where the reference standard is itself symptom-based. A supervised classifier trained to reproduce a Rome IV diagnosis can at best approximate that criterion - including its instability across criteria versions - but cannot validate it. Conversely, unsupervised stratification of microbiome or symptom data may identify biologically coherent groups whose clinical meaning is unproven, and work in DGBI has highlighted the sensitivity of such subgroupings to methodological choice[14]. Most published IBS applications remain supervised classification of cross-sectional data. DL, though central to AI in luminal endoscopy[9] and to medical image classification more broadly[15], has had little application to functional gastrointestinal disease; and while NLP has been used to extract patient-reported outcomes from inflammatory bowel disease records[16], we identified no comparable work in IBS.
AI APPLICATIONS IN IBS DIAGNOSIS AND SUBTYPING
Diagnostic classification models
Differentiating IBS from organic gastrointestinal disease, and identifying patients who require invasive investigation, are central clinical challenges. Several studies have applied machine-learning algorithms to develop diagnostic classifiers using combinations of clinical features, laboratory parameters, and questionnaire responses. Table 1 summarizes representative AI studies across diagnosis, subtyping, and treatment prediction, together with their cohorts, performance, and principal limitations. Random forest models built on routine blood, urine and faecal laboratory markers have separated inflammatory bowel disease from non-inflammatory bowel disease controls, with mean average precision of 97% for Crohn’s disease and 91% for ulcerative colitis; that discovery cohort included no IBS patients, however, so this constitutes indirect evidence only, and its applicability to the clinically harder task of distinguishing IBS from organic disease remains unproven[17]. These results should be interpreted cautiously: A 2025 systematic review of AI models in IBS identified only six eligible studies (1366 participants) and found the datasets, biomarkers, and validation methods too heterogeneous to support meta-analysis, underscoring the early-stage nature of this evidence[18].
In the broader context of functional GI symptoms, Maslekar et al[19] developed an artificial neural network for patients presenting to colorectal clinics, integrating demographic data, symptom characteristics, and basic laboratory results to predict the presence of significant pathology; the model showed high discriminative performance for identifying patients likely to harbor organic disease, suggesting a potential role in prioritizing colonoscopy while maintaining diagnostic safety. The predictive accuracy of the neural network was 95% in the training group and 90% (95%CI: 84-96) in the internal validation set, compared with 75% for clinical assessment, and 89% in a separate external validation set of 50 patients; the cohort comprised unselected colorectal clinic attendees rather than an IBS population, so this too is indirect evidence[19].
IBS subtype differentiation
Traditional IBS subtyping relies on the Bristol Stool Form Scale and predominant bowel-habit patterns[20]; this approach is widely considered subject to intra-individual variability and subjective interpretation, although that limitation is asserted rather than formally quantified in the source cited. Machine-learning approaches offer potentially more robust classification by integrating multiple data modalities. Tap et al[21] used a supervised discriminant procedure to derive a 90-operational-taxonomic-unit faecal and mucosal microbiota signature that discriminated severe from mild-to-moderate IBS and tracked symptom severity rather than bowel-habit subtype.
Multi-modal analytical frameworks that combine patient-reported symptom severity, psychological assessments, and physiological measurements have been proposed as a route to identifying biologically distinct IBS endotypes with differential treatment responses. Such integrated approaches illustrate the potential for AI-assisted phenotyping that extends beyond traditional bowel-habit categories, though this remains largely conceptual and requires prospective validation.
Clinical implementation perspectives
Most AI applications in IBS remain investigational, but early implementations are instructive. Several academic centers have begun exploring machine-learning approaches to support clinical decision-making, including efforts to refine IBS subtyping and personalize treatment. Recent work has evaluated the diagnostic performance of AI models across IBS populations; the studies identified to date have all used supervised classifiers rather than unsupervised subgrouping[18]. Early evidence suggests these methods may help clinicians identify symptom clusters more efficiently and guide individualized management, though comparative effectiveness against standard care is not yet established.
Digital therapeutics such as Zemedy and Mahana deliver structured cognitive behavioral therapy (CBT) for IBS through automated, rule-based programs; neither has published evidence of a machine-learning component, and they are therefore best regarded as adjacent to, rather than examples of, AI-driven care[22]. Real-world data nonetheless illustrate the implementation challenge: Among 843 patients prescribed a digital CBT program, only 38% completed half the available sessions, although those who did showed significant symptom reduction[23]. Independent validation of any underlying algorithm is lacking.
GUT MICROBIOME ANALYSIS THROUGH AI
Microbiome profiling in IBS
The gut microbiome contributes to IBS pathophysiology, with multiple studies documenting reduced Bifidobacterium and Faecalibacterium prausnitzii abundance in IBS patients, alongside inconsistent findings for overall microbial diversity, which was variously reported as decreased or unchanged[24]. Translating these findings into clinically actionable insights has proven difficult, owing to substantial inter-individual variability and technical heterogeneity in sequencing protocols.
Machine-learning algorithms, particularly random forests, can identify microbiome and metabolome signatures associated with IBS status; in one cohort, random forest models separated IBS from controls with an area under the curve of 0.81 for faecal microbiota and 0.69 for the faecal metabolome[25]. These models have not been benchmarked head-to-head against conventional statistical approaches, so claims of superiority remain unsubstantiated. These approaches accommodate the high-dimensional, compositional nature of microbiome data and can capture complex non-linear relationships between microbial taxa and clinical phenotypes.
Microbial functional profiling
Beyond taxonomic composition, the functional potential of the gut microbiome - assessed through 16S rRNA profiling, metatranscriptomics and untargeted metabolomics - provides mechanistic insight into IBS[26]. Integrative analyses have implicated disrupted short-chain fatty acid production and bile acid metabolism, while microbial tryptophan metabolism has been characterized as a distinct route linking the microbiota to host signalling[27].
Vervier et al[28] combined shotgun metagenomics with unsupervised analysis and identified two microbiota subtypes among IBS patients - a “pathogenic-like” profile (enriched in Firmicutes and amino-acid/carbohydrate metabolism genes and depleted in Bacteroidetes) and a “health-like” profile resembling that of controls. Critically, these subtypes showed distinct responses to the low-FODMAP diet: The pathogenic-like signature shifted toward a health-associated composition and its clinical response was greater, whereas the health-like and control microbiota were largely unaffected[28]. This work exemplifies how computational stratification of microbiome data can identify subgroups with differential therapeutic response.
Predictive models for microbiome-based therapies
Probiotics, prebiotics, synbiotics and antibiotics show variable efficacy in IBS, the clearest pooled benefit being for rifaximin in non-constipated disease[29]; faecal microbiota transplantation (FMT) falls outside that synthesis and its efficacy in IBS remains contested. AI-based approaches have been proposed to predict individual response from baseline microbiome configuration. Longitudinal multi-omics work by Mars et al[10] did not itself provide a treatment-prediction classifier, but by integrating microbiome, metabolome, and host data over time it identified IBS subset-specific mechanisms - including purine metabolism as a novel host-microbial pathway - that could inform future response-prediction models and microbiome-directed targets[10].
Computational approaches have also been proposed to optimize FMT donor selection by matching donor and recipient microbiome characteristics, potentially improving engraftment and outcomes, though robust prospective evidence is still lacking[30].
INTEGRATION OF PATIENT-REPORTED OUTCOMES
Digital symptom monitoring
Traditional symptom assessment in IBS relies on retrospective recall during clinical encounters. In 273 patients with Rome III IBS, retrospective seven-day recall systematically overestimated average pain intensity, urgency and disordered stool relative to contemporaneous end-of-day diary entries, with roughly one-fifth to one-half of patients deviating by two or more scale points, depending on the symptom[31]. Smartphone-based platforms allow real-time symptom and dietary logging: A purpose-built IBS food and gastrointestinal symptom journal application recorded four symptoms four times daily alongside all meals, and proved feasible and acceptable in a small pilot[32]. Such tools generate longitudinal datasets that are in principle amenable to AI analysis, and both recurrent neural network[33] and transformer-based[34] architectures have been applied to longitudinal clinical records in other settings. We identified no peer-reviewed study, however, applying ML to longitudinal patient-collected IBS symptom or dietary data to identify triggers or predict flares; this remains an open research gap rather than an established capability.
Quality-of-life assessment
IBS substantially diminishes health-related quality of life across physical, psychological, and social domains, as documented by disease-specific instruments such as the IBS-quality of life measure questionnaire[35]. Beyond core symptoms, patients experience secondary effects - fatigue, dietary restriction, unpredictable exacerbations, heightened anxiety, social avoidance, and reduced self-efficacy - that collectively compromise daily functioning. While validated instruments enable standardized assessment, their fixed-item formats may not capture the nuanced, highly variable experiences characteristic of IBS - an interpretation we advance here rather than one tested in the instrument's validation studies.
NLP offers a complementary approach by enabling systematic analysis of unstructured text from patient narratives, open-ended questionnaire responses, online health forums, patient-physician communications, and social media. Rather than constraining patients to predetermined response categories, NLP methods - including topic modeling and relationship-extraction algorithms - can surface emergent themes and previously unrecognized symptom patterns that may not have been anticipated during instrument development[36]; we identified no study applying them to coping strategies or quality-of-life dimensions in IBS. A review by Gonzalez-Hernandez et al[36] traces the shift from rule-based to learning-based analysis of patient-generated text, while emphasising persistent obstacles including scarce annotated data, noisy sources, and the difficulty of mapping lay expression onto standard medical concepts[36].
Sentiment features derived from patient communications have been used as predictors in clinical text analysis, for example in classifying health literacy from secure messages[37]; we identified no study using them to quantify emotional valence as a clinical measure in IBS. In IBS, this can help distinguish emotional distress - expressions of anxiety, frustration, or hopelessness - from symptom severity itself, a clinically meaningful distinction because distress may persist despite symptom control, or vice versa. Sentiment analysis can also gauge treatment satisfaction by identifying which aspects of therapy generate positive or negative responses, informing shared decision-making. Applied to routine communications such as patient-portal messages, these methods could provide ongoing assessment between formal encounters, and, as complementary trial endpoints, may detect treatment effects on emotional well-being that primary symptom endpoints miss. These remain proposals rather than demonstrated applications in IBS.
Psychological comorbidity detection
Anxiety and depressive symptom levels are markedly higher in IBS populations, with meta-analysis reporting standardized mean differences of 0.76 and 0.80 respectively relative to healthy controls[38]. This reflects shared neurobiological vulnerabilities, bidirectional gut-brain signaling, and the psychosocial impact of chronic, unpredictable symptoms. The authors note that psychological factors are widely regarded as moderators of symptom severity and treatment response, although their meta-analysis assessed only cross-sectional anxiety and depression levels relative to healthy controls and did not itself test clinical trajectory or treatment response[38]. Conversely, a randomized controlled trial by Lackner et al[39] showed that IBS-specific CBT improved gastrointestinal symptom severity relative to an education control, with 55%-61% of patients rated as responders immediately after treatment compared with 43% of education controls, although between-group differences had narrowed by six-month follow-up; the trial included no pharmacological comparator, so relative efficacy against drug therapy cannot be inferred from it[39].
Machine-learning methods can advance psychological-comorbidity detection by integrating screening scores, symptom profiles, and gut-brain axis biomarkers into unified predictive models that capture non-linear relationships traditional statistics may miss. For example, elevated anxiety scores combined with pain-predominant, stress-triggered symptoms and biomarker patterns (such as altered inflammatory or cortisol profiles) might identify patients most likely to benefit from psychological intervention. Objective gut-brain axis correlates - autonomic function, inflammatory cytokines, hypothalamic-pituitary-adrenal-axis reactivity, and potentially microbiome composition - can be weighted alongside self-report to stratify patients into management pathways, flag inadequate responders who warrant psychological assessment, and support personalized, shared decision-making. Framing psychological factors as legitimate pathophysiological components, rather than dismissing symptoms as “merely psychological”, validates patient experience while expanding therapeutic options. It should be emphasized that these remain conceptual proposals: We identified no published study applying ML to the detection or stratification of psychological comorbidity specifically in IBS populations, and we therefore flag this as a research gap rather than an existing application.
TREATMENT OPTIMIZATION AND PERSONALIZED MEDICINE
Dietary intervention prediction
The low-FODMAP diet is a widely used dietary therapy for IBS, yet a substantial minority of patients do not respond: In a randomized controlled trial, 57% achieved adequate symptom relief on a low-FODMAP diet compared with 38% on a sham diet, leaving 43% without adequate benefit[40]. Baseline microbial profiling has been explored as a means of identifying likely responders in advance. Bennet et al[41] reported that faecal bacterial profiles discriminated responders from non-responders among 61 analysed patients, with non-responders showing a higher baseline dysbiosis index, and Chumpitazi et al[42] found that paediatric responders were enriched at baseline for saccharolytic taxa. Neither study reported discrimination metrics such as sensitivity, specificity or area under the curve, neither included an external validation cohort, and their directions of effect are not concordant. These findings should therefore be regarded as hypothesis-generating associations at high risk of overfitting, and not as clinically actionable predictive thresholds.
Disease-trajectory prediction
Predicting disease trajectory and treatment responsiveness remains challenging in IBS. Gradient-boosting and other ensemble methods are well established for risk prediction in clinical medicine generally[43], and could in principle incorporate baseline symptom severity, psychological comorbidity, dietary patterns and early treatment response to generate individualized risk scores. We identified no study applying these methods to longitudinal IBS cohorts to forecast symptom exacerbations or quality-of-life deterioration; this is therefore a proposed application rather than demonstrated evidence.
Pharmacological treatment selection
Current pharmacological options for IBS include antispasmodics, laxatives, antidiarrheals, antidepressants, and peripherally acting agents targeting serotonin and guanylate cyclase pathways[44]. Treatment selection remains largely empirical, with limited ability to predict individual drug response. A second-generation artificial-intelligence platform incorporating therapeutic variability has been proposed for improving treatment response in IBS, although that work is a concept and review paper reporting no original data and no pharmacogenomic component[45]; machine-learning approaches integrating clinical phenotype with pharmacogenomic data remain untested in IBS.
Psychological phenotype is increasingly recognized as relevant to treatment matching[46], and mixture-model analysis has identified IBS subgroups defined by distinct gastrointestinal, extraintestinal somatic, and psychological symptom profiles[47]. Whether these subgroups respond differentially to neuromodulators vs gut-directed pharmacotherapy remains a hypothesis that has not been tested prospectively. Reinforcement-learning approaches raise the possibility of adaptive protocols that adjust therapy based on ongoing symptom monitoring; this has been explored retrospectively in critical care using observational intensive-care datasets, but has not been evaluated prospectively in patients and has not been applied in IBS, so it remains a proposed application here[48].
Multi-modal treatment algorithms
IBS management often requires multi-component interventions addressing dietary, pharmacological, psychological, and lifestyle factors[49]. Figure 2 outlines a proposed AI-assisted clinical pathway from triage through personalized management and closed-loop monitoring. Clinical decision-support systems have been advocated in other domains of medicine as a means of translating model outputs into actionable recommendations[50]. Analogous AI-driven management algorithms for IBS remain at a very early stage. The only comparator-controlled evidence we identified is a small open-label pilot in 25 patients with IBS-M, in which a microbiome-based, AI-personalized diet produced a greater fall in symptom severity than a standard IBS diet; allocation was by matching rather than randomization, the primary outcome was unblinded and subjective, and follow-up was six weeks[51]. This is hypothesis-generating only, and adequately powered randomized trials are required.
Figure 2 Proposed artificial intelligence-assisted clinical pathway for irritable bowel syndrome, extending from initial triage and diagnostic classification through personalized, multi-modal management (dietary, pharmacological, and psychological interventions) to closed-loop symptom monitoring and adaptive adjustment of therapy.
This pathway is hypothetical and is presented as a research framework, not as a validated or guideline-endorsed clinical algorithm. No artificial intelligence classifier is currently validated for routine IBS diagnosis, endoscopy triage, endotype classification or treatment selection, and none of the steps shown should replace standard clinical assessment, alarm-feature evaluation or guideline-based investigation. AI: Artificial intelligence; IBS: Irritable bowel syndrome; IBS-C: Irritable bowel syndrome with predominant constipation; IBS-D: Irritable bowel syndrome with predominant diarrhea; IBS-M: Irritable bowel syndrome with mixed bowel habits; IBS-U: Irritable bowel syndrome with unclassified.
CHALLENGES AND LIMITATIONS
The principal barriers to clinical translation, and corresponding mitigation strategies, are summarized in Table 3.
Table 3 Key challenges to clinical translation of artificial intelligence in irritable bowel syndrome and corresponding mitigation strategies.
Challenge
Why it matters for IBS
Mitigation strategies
Data standardization
Heterogeneous diagnostic criteria, symptom instruments, and sequencing protocols limit generalizability
Rome Foundation frameworks; harmonized data-collection standards; shared microbiome pipelines
Model interpretability
“Black-box” predictions are hard for clinicians to trust or act on
Explainable AI (SHAP values, attention visualization); prefer interpretable models for high-stakes decisions
External validation
Most models are single-center; performance may not transfer across settings
Robust AI models require large, high-quality datasets collected under standardized protocols. Transparent documentation of a model’s intended use, validation and limitations has been proposed as a means of supporting appropriate interpretation by the clinicians expected to rely on it[52]. Current IBS research is limited by heterogeneous diagnostic criteria, variable symptom-assessment instruments, and inconsistent microbiome-sequencing methodologies, all of which constrain generalizability. Initiatives such as the Rome Foundation Global Epidemiology Study are working toward standardized epidemiological frameworks[3].
Model interpretability
Complex deep-learning models often operate as “black boxes” with limited interpretability[53]. Clinicians require transparent, explainable predictions to trust AI recommendations and integrate them into practice[54]. Explainable-AI techniques, including SHapley Additive exPlanations values, offer promise for elucidating model decision-making[55].
External validation
Many AI models in IBS have been developed and validated within single-center cohorts, raising concerns about generalizability across populations and settings, mirroring problems documented across medical AI more broadly[56]. Rigorous external validation using independent datasets from multiple regions is essential before clinical implementation, as is temporal validation to ensure performance stability given evolving diagnostic criteria and treatments[57]; AI-specific reporting guidance additionally asks that performance be reported across sociodemographic and ethnic subgroups[58]. Reporting and appraisal standards have since been updated specifically for AI-based prediction models: Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis + AI defines reporting requirements for models developed using regression or ML[58], PROBAST+AI provides a corresponding risk-of-bias and applicability assessment[59], and DECIDE-AI addresses the early-stage clinical evaluation of AI-driven decision-support systems[60]. Future IBS models should be developed and reported against these frameworks.
Privacy, consent, and algorithmic bias
Collecting and analyzing sensitive health data, including microbiome profiles and psychological assessments, raises important privacy and consent considerations[61]. Regulatory frameworks such as General Data Protection Regulation and Health Insurance Portability and Accountability Act impose strict requirements on data handling and consent, which can constrain data sharing and collaborative model development[62]. Algorithmic bias may perpetuate health disparities through several mechanisms, including non-representative training data and biased outcome proxies - as demonstrated where health-care cost was used as a surrogate for health need in a large single-centre cohort of insured primary care patients[63] - requiring deliberate attention to fairness and equity.
Cost-effectiveness
Although IBS imposes substantial per-patient direct and indirect costs[4], the cost-effectiveness of AI-based interventions is uncertain. Development and maintenance costs - data infrastructure, computation, and ongoing model updating - are substantial. Healthcare systems require evidence that these investments reduce diagnostic delays, unnecessary procedures, and inefficient treatment. It has been argued that AI may improve the value of colonoscopy by reducing unnecessary polypectomy and histopathology through optical diagnosis of diminutive lesions[64]; we identified no formal economic modelling in IBS, and comprehensive analyses incorporating all stakeholder perspectives are needed.
Regulatory landscape
AI-based clinical decision support and digital therapeutics face an evolving regulatory environment. In the United States, the Food and Drug Administration regulates Software as a Medical Device (SaMD), with AI/ML tools classified by clinical risk; its 2021 AI/ML-based SaMD Action Plan emphasizes predetermined change-control plans that accommodate algorithm updates while preserving safety. Under this framework most Class II devices are subject to 510 (k) review and Class III devices to premarket approval[65]. No classification determination has been published for an IBS-specific tool; a novel tool without a suitable predicate device would be more likely to proceed through the de novo pathway, and classification differs between jurisdictions. Uncertainty persists regarding continuously learning algorithms and the harmonization of standards across jurisdictions, both of which may slow translation[65].
Patient acceptance and trust
Beyond data privacy, patient acceptance of and trust in AI recommendations remain understudied. Surveys suggest patients have concerns about algorithmic bias, loss of human connection in care, and transparency in how conclusions are reached - factors that will shape real-world adoption and warrant dedicated study.
Methodological risks in AI prediction studies
Beyond the barriers above, the small IBS AI literature is exposed to a set of methodological failure modes that are easily overlooked when performance metrics are reported in isolation. Overfitting is the most immediate: With cohorts of a few dozen to a few hundred participants and feature spaces running to hundreds of microbial taxa or laboratory variables, a model can achieve excellent apparent discrimination while learning cohort-specific noise. Data leakage - for example, feature selection or normalization performed before rather than within cross-validation folds - inflates apparent performance further and is frequently under-reported. Spectrum bias is a particular concern in IBS: A model trained on tertiary-referral patients with established diagnoses faces a very different case mix from primary care, where the discrimination that matters is between IBS and undifferentiated abdominal symptoms.
Calibration is almost never reported. Discrimination indicates whether a model ranks patients correctly; calibration indicates whether its predicted probabilities correspond to observed risks, and it is calibration that determines whether a predicted 70% probability of low-FODMAP response can be acted upon. Dataset shift compounds this, since sequencing protocols, symptom instruments and diagnostic criteria all change over time, so that a model developed under Rome III may behave differently under Rome IV. Discrimination and calibration together establish only technical performance, not clinical utility: Demonstrating that a model changes management in a way that benefits patients requires prospective comparative studies, of which none exists in IBS.
These risks carry a specific clinical hazard in the triage setting. A model that reduces colonoscopy in low-risk patients will, at some threshold, miss organic disease, and inappropriate reassurance in a patient with undiagnosed inflammatory bowel disease or malignancy is a materially different harm from an unnecessary procedure. Any AI-assisted triage pathway therefore requires explicit consideration of the asymmetry between false-positive and false-negative costs, pre-specified safety thresholds, and retention of alarm-feature assessment and guideline-based investigation as the clinical floor.
Microbiome-specific sources of irreproducibility
Microbiome-based models carry reproducibility problems specific to the data type, which compound the general risks above. Sequencing data are compositional: The instrument imposes an arbitrary total, so taxon abundances are relative rather than absolute, and applying conventional correlation or differential-abundance methods to such data produces artefacts that are not biological[66]. Technical variation is equally consequential. In the Microbiome Quality Control study, blinded specimen sets were sequenced by 15 laboratories and processed through nine bioinformatic protocols; variability depended most on biospecimen type and origin, followed by DNA extraction, sample-handling environment and bioinformatic processing[67]. A model trained in one cohort therefore encounters not only a different case mix but a different measurement process.
Host and clinical variables are a third source. In an analysis of population-scale cohorts, alcohol intake and bowel-movement quality emerged as unexpectedly strong sources of gut microbiota variance that were distributed differently between cases and controls; after cases and controls were matched on such variables, the 26 taxa that had distinguished type 2 diabetes fell to none, and machine-learning classification accuracy declined substantially[68]. The implication for IBS is direct, because stool form is simultaneously a diagnostic criterion, a determinant of microbiota composition and a candidate model feature - a configuration in which a classifier may learn bowel habit rather than disease. Diet and medication exposure act similarly. These considerations do not invalidate microbiome-based prediction in IBS, but they mean that reported discrimination should be treated as provisional until a model is shown to survive compositionally aware analysis, protocol-independent replication and adjustment for host confounders. None of the IBS studies reviewed here reports all three.
Implementation across regions and resource settings
The prerequisites for the applications described here are unevenly distributed. Symptom-monitoring and digital-therapeutic approaches assume reliable smartphone access, data connectivity and literacy in the language of the interface. Microbiome-guided strategies assume sequencing at a cost and turnaround compatible with clinical decisions. Treatment-selection models assume the availability of the interventions they recommend, including specialist gastroenterology review and the dietetic supervision required for restrictive diets such as the low-FODMAP approach, where structured reintroduction is needed to avoid unnecessary long-term restriction. Where those inputs are absent, a model that performs well in a well-resourced centre may be unusable, or may recommend care that cannot be delivered.
This bears on equity as well as feasibility. The IBS AI literature originates in a small number of countries - the eligible studies in the available systematic review came from the United Kingdom, China and Japan[18] - and the wider clinical prediction literature shows a comparable concentration[56]. Populations in which these models were developed therefore differ in diet, ethnicity and background microbiota from most patients with IBS worldwide. Deployment requires local validation rather than assumed transportability, and cost-effectiveness analysis framed around the resources of the implementing health system rather than the developing one. In resource-constrained settings, the more realistic near-term role is triage support that reduces avoidable investigation, rather than microbiome- or omics-based personalization.
SYNTHESIS OF THE EVIDENCE
Taken together, the evidence base for AI in IBS is early-stage, and its strongest components are of the weakest design. No randomized trial has evaluated an AI-guided management strategy against standard care, and no prospective study has tested whether any model changes clinical decisions or patient outcomes. The available work consists of cross-sectional diagnostic and association analyses, most with internal validation only; the sole comparator-controlled intervention we identified was an open-label, non-randomized pilot of 25 patients in which allocation was by matching rather than randomization[51]. On any conventional hierarchy of evidence, this places the field below the threshold at which clinical recommendations can be made.
Consistency cannot presently be assessed, because few findings have been tested twice. The available systematic review identified six eligible studies whose datasets, biomarkers and validation methods were too heterogeneous to pool, with reported diagnostic accuracies spanning 54% to 98%[18]. Where two studies do address the same question, they do not always agree: Baseline microbial predictors of low-FODMAP response differ in direction between the adult and paediatric cohorts summarized in Table 1. Directness is similarly limited. Of the studies in Table 1, only a minority were conducted in IBS populations with an IBS-relevant outcome; the remainder extrapolate from inflammatory bowel disease, unselected colorectal-clinic cohorts, critical care or generic methodological work. We have labelled these grades explicitly rather than presenting them as a single body of evidence, because the distinction determines what may be concluded from them.
Clinical applicability is correspondingly narrow. The most plausible near-term contribution is decision support at triage - helping to identify patients in whom organic disease is unlikely - because that is where the strongest indirect evidence lies and where the clinical question is best posed. Even there, the asymmetry between missing organic disease and performing an unnecessary procedure means such a tool must be additive to, rather than a substitute for, alarm-feature assessment and guideline-based investigation. Microbiome-guided therapy selection, psychological-comorbidity detection and closed-loop symptom management remain conceptually attractive but, on present evidence, unproven in IBS. The practical consequence is that the immediate priority for the field is not further model development but replication, external validation and prospective evaluation of the models that already exist.
FUTURE DIRECTIONS AND RESEARCH PRIORITIES
The priorities below are framed as testable questions rather than general aspirations. Table 4 sets out the minimum design and reporting standards we suggest should apply to any future AI study in IBS, covering intended use, reference standard, data elements, external and temporal validation, calibration, fairness, prospective impact assessment and reporting.
Table 4 Proposed minimum design and reporting standards for future artificial intelligence studies in irritable bowel syndrome.
Domain
Minimum standard proposed
Why it matters in IBS
Intended use
State the clinical question, target population, care setting and the decision the model informs; specify whether it is diagnostic, prognostic or predictive of treatment response
Most IBS models report discrimination without stating the decision they are intended to support
Reference standard
Define how labels were assigned (Rome IV criteria with documented exclusion of organic disease), who applied them and whether assessment was blinded
IBS has no biomarker, so label uncertainty is inherited directly by any supervised model
Data elements
Pre-specify candidate features and their provenance; report missing data and how it was handled; for microbiome data report extraction, sequencing and bioinformatic pipeline
Technical factors are a dominant source of microbiome variance and differ between centres
External validation
Evaluate in at least one cohort independent of the development cohort in geography and recruitment pathway
Reported IBS models are overwhelmingly validated internally only
Temporal validation
Test stability over time and across changes in diagnostic criteria, symptom instruments and sequencing protocols
The shift from Rome III to Rome IV alters both prevalence and case mix
Calibration
Report calibration slope and intercept, or a calibration plot, alongside discrimination
Calibration determines whether a predicted probability of treatment response can be acted upon; it is almost never reported
Fairness
Report performance across sex, age, ethnicity and socioeconomic strata, and state explicitly any outcome proxy used in place of the outcome of interest
The evidence base is geographically narrow and IBS cohorts are predominantly female
Prospective impact
Before adoption, evaluate the effect on clinical decisions, resource use and patient outcomes in a prospective comparative design
No prospective impact study has been undertaken in IBS
Reporting standards
Report to TRIPOD+AI[58]; appraise risk of bias with PROBAST+AI[59]; follow DECIDE-AI[60] for early clinical evaluation
Aligns IBS work with current reporting norms for AI prediction models
The rationale for integrating genomic, transcriptomic, proteomic, metabolomic and microbiome data is well established[69], and network-based and multi-view integration methods have been developed for this purpose in biology and medicine[70]; their extension to graph neural networks, and to the gut-brain-microbiome axis specifically, remains prospective. The testable question is whether a multi-omics classifier outperforms a model built on clinical and symptom variables alone for IBS subtype assignment or treatment-response prediction, evaluated in the same cohort with external validation.
Real-time clinical decision support
Integration with electronic health records, coupled with interfaces that fit clinical workflow and require little additional training, has been argued to be a precondition for adoption[71]. Federated learning, which trains models across institutions without sharing raw data, may supply the multi-centre datasets this field lacks while preserving privacy[72]. The testable question is whether a federated model trained across several centres retains its discrimination and calibration at each participating site.
Predictive biomarker discovery
ML has been applied to high-dimensional omics data to identify candidate biomarkers in other disease areas[73]; in IBS, plausible candidates include microbial metabolites, host immune markers and gut-brain axis mediators, although none has yet been validated for diagnostic classification or treatment-response prediction. The testable question is whether a candidate biomarker identified in a discovery cohort replicates in an independent cohort against a pre-specified threshold fixed in advance.
Digital therapeutics and closed-loop systems
Mobile health interventions can support adherence in chronic disease, although a systematic review found their effects on clinical outcomes to be mixed[74]. The convergence of AI, wearable sensors and digital therapeutics raises the further prospect of closed-loop systems that continuously monitor symptoms and adjust interventions; we identified no such system in IBS. The testable question is whether an adaptive, algorithm-guided intervention improves symptom severity relative to the same intervention delivered on a fixed protocol.
CONCLUSION
AI has meaningful potential to refine the diagnosis and management of IBS and FGIDs. By integrating patient-reported outcomes, clinical features, microbiome data, and multi-omics profiles, AI approaches offer new capabilities for diagnostic refinement, mechanistic discovery, and treatment personalization. Early studies are consistent with a role in differentiating IBS from organic pathology, identifying candidate subgroups and predicting response to the low-FODMAP diet; each of these findings, however, rests on small and mostly internally validated cohorts, and none has been replicated prospectively.
These capabilities, however, remain early-stage. The current evidence base is small, heterogeneous, and largely confined to single-center cohorts, and few models have undergone external validation. Standardized data collection, rigorous external and temporal validation, improved interpretability, and careful attention to privacy, equity, and regulation are prerequisites for clinical translation. User-friendly decision-support systems integrated with existing infrastructure will be essential for clinician adoption.
Future research should prioritize multi-omics integration, real-time symptom-monitoring platforms, and adaptive treatment algorithms, supported by collaborative, multi-institutional datasets. Whether AI ultimately reduces diagnostic delay, limits ineffective treatment or improves outcomes in IBS is at present an open question rather than a demonstrated benefit, since no prospective implementation study has tested it. AI tools in IBS should accordingly be regarded as investigational and adjunctive, complementing rather than substituting for guideline-based clinical assessment, until external validation and prospective evaluation establish otherwise.
Sperber AD, Bangdiwala SI, Drossman DA, Ghoshal UC, Simren M, Tack J, Whitehead WE, Dumitrascu DL, Fang X, Fukudo S, Kellow J, Okeke E, Quigley EMM, Schmulson M, Whorwell P, Archampong T, Adibi P, Andresen V, Benninga MA, Bonaz B, Bor S, Fernandez LB, Choi SC, Corazziari ES, Francisconi C, Hani A, Lazebnik L, Lee YY, Mulak A, Rahman MM, Santos J, Setshedi M, Syam AF, Vanner S, Wong RK, Lopez-Colombo A, Costa V, Dickman R, Kanazawa M, Keshteli AH, Khatun R, Maleki I, Poitras P, Pratap N, Stefanyuk O, Thomson S, Zeevenhooven J, Palsson OS. Worldwide Prevalence and Burden of Functional Gastrointestinal Disorders, Results of Rome Foundation Global Study.Gastroenterology. 2021;160:99-114.e3.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 1768][Cited by in RCA: 1592][Article Influence: 318.4][Reference Citation Analysis (15)]
Mars RAT, Yang Y, Ward T, Houtti M, Priya S, Lekatz HR, Tang X, Sun Z, Kalari KR, Korem T, Bhattarai Y, Zheng T, Bar N, Frost G, Johnson AJ, van Treuren W, Han S, Ordog T, Grover M, Sonnenburg J, D'Amato M, Camilleri M, Elinav E, Segal E, Blekhman R, Farrugia G, Swann JR, Knights D, Kashyap PC. Longitudinal Multi-omics Reveals Subset-Specific Mechanisms Underlying Irritable Bowel Syndrome.Cell. 2020;182:1460-1473.e17.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 360][Cited by in RCA: 325][Article Influence: 54.2][Reference Citation Analysis (1)]
Vervier K, Moss S, Kumar N, Adoum A, Barne M, Browne H, Kaser A, Kiely CJ, Neville BA, Powell N, Raine T, Stares MD, Zhu A, De La Revilla Negro J, Lawley TD, Parkes M. Two microbiota subtypes identified in irritable bowel syndrome with distinct responses to the low FODMAP diet.Gut. 2022;71:1821-1830.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 130][Cited by in RCA: 115][Article Influence: 28.8][Reference Citation Analysis (0)]
Su Q, Chen S, Lau LH, Lui RN, Wang Y, Xu Z, Cheung CP, Ching JYL, Shen X, Peng Y, Tun HM, Ianiro G, Rubin D, Chang EB, Chan FKL, Ng SC. Artificial intelligence-driven donor-recipient gut microbiome matching for optimized fecal microbiota transplantation.Cell Rep. 2026;45:117301.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 4][Reference Citation Analysis (0)]
Choi E, Bahadori MT, Schuetz A, Stewart WF, Sun J. Doctor AI: Predicting Clinical Events via Recurrent Neural Networks.JMLR Workshop Conf Proc. 2016;56:301-318.
[PubMed] [DOI]
Staudacher HM, Lomer MCE, Farquharson FM, Louis P, Fava F, Franciosi E, Scholz M, Tuohy KM, Lindsay JO, Irving PM, Whelan K. A Diet Low in FODMAPs Reduces Symptoms in Patients With Irritable Bowel Syndrome and A Probiotic Restores Bifidobacterium Species: A Randomized Controlled Trial.Gastroenterology. 2017;153:936-947.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 382][Cited by in RCA: 325][Article Influence: 36.1][Reference Citation Analysis (2)]
Hershko-Moshe A, Hasin Y, Nevo-Shor A, Etzion O, Ilan Y. Improving the Treatment Response of Patients with Irritable Bowel Syndrome: Implementing a Second-Generation Artificial Intelligence System for Overcoming Resistance.GastroHep. 2022;2022:1-12.
[PubMed] [DOI] [Full Text]
Vasant DH, Paine PA, Black CJ, Houghton LA, Everitt HA, Corsetti M, Agrawal A, Aziz I, Farmer AD, Eugenicos MP, Moss-Morris R, Yiannakou Y, Ford AC. British Society of Gastroenterology guidelines on the management of irritable bowel syndrome.Gut. 2021;70:1214-1240.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 423][Cited by in RCA: 388][Article Influence: 77.6][Reference Citation Analysis (3)]
Tonekaboni S, Joshi S, McCradden MD, Goldenberg A.
What clinicians want: contextualizing explainable machine learning for clinical end use. In: Proceedings of the 4th Machine Learning for Healthcare Conference. PMLR 2019; 106: 359-380.
[PubMed] [DOI]
Vasey B, Nagendran M, Campbell B, Clifton DA, Collins GS, Denaxas S, Denniston AK, Faes L, Geerts B, Ibrahim M, Liu X, Mateen BA, Mathur P, McCradden MD, Morgan L, Ordish J, Rogers C, Saria S, Ting DSW, Watkinson P, Weber W, Wheatstone P, McCulloch P; DECIDE-AI expert group. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI.Nat Med. 2022;28:924-933.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 372][Cited by in RCA: 488][Article Influence: 122.0][Reference Citation Analysis (5)]
Sinha R, Abu-Ali G, Vogtmann E, Fodor AA, Ren B, Amir A, Schwager E, Crabtree J, Ma S; Microbiome Quality Control Project Consortium, Abnet CC, Knight R, White O, Huttenhower C. Assessment of variation in microbial community amplicon sequencing by the Microbiome Quality Control (MBQC) project consortium.Nat Biotechnol. 2017;35:1077-1086.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 280][Cited by in RCA: 406][Article Influence: 45.1][Reference Citation Analysis (0)]
