Role of artificial intelligence in irritable bowel syndrome: A narrative review

Table 1 Summary of representative artificial intelligence studies in irritable bowel syndrome, with evidence grade

Ref.	AI method and task	Population and sample size	Primary outcome definition	Validation approach	Reported performance and uncertainty	Events or responders reported	Evidence grade	Key limitation
Maslekar et al[19], 2010	Artificial neural network; predict organic pathology to prioritise colonoscopy	300 unselected colorectal-clinic attendees; no IBS cohort	Positive finding at lower gastrointestinal endoscopy (polyp, cancer, diverticular disease or colitis)	Internal validation plus external set of 50 patients	95% training; 90% internal validation (95%CI: 84-96) vs 75% clinical assessment; 89% external	Not reported	Indirect	Non-IBS population; predates modern validation standards
Tap et al[21], 2017	Supervised discriminant analysis; microbiota signature of symptom severity	IBS patients and controls; faecal and mucosal microbiota	IBS symptom severity category (severe vs mild-to-moderate)	Internal only	90-OTU signature tracking severity, independent of bowel-habit subtype; no discrimination metric reported	Not reported	Direct	Association only; not a diagnostic model
Vervier et al[28], 2022	Unsupervised metagenomic stratification; subtypes and low-FODMAP response	IBS patients; shotgun metagenomics	Metagenomic subtype assignment (IBSP/IBSH) and symptom response to low-FODMAP diet	Internal only	Greater FODMAP response in IBSP (P = 0.02); no AUC or confidence interval reported	Not reported	Direct	Small cohort; requires external validation
Mars et al[10], 2020	Longitudinal multi-omics integration; subset-specific mechanisms	IBS-C, IBS-D and controls; serial multi-omics	Not applicable; mechanistic analysis with no prediction target	Not applicable	Purine metabolism identified as a host-microbial pathway; no predictive metric	Not applicable	Direct	Mechanistic; no predictive classifier derived
Kraszewski et al[17], 2021	Random forest; classify IBD from routine laboratory markers	702 records (Crohn’s, ulcerative colitis, non-inflammatory controls); no IBS patients	Crohn’s disease or ulcerative colitis vs non-inflammatory control status	Internal discovery cohort only	Mean average precision 97% (Crohn’s), 91% (ulcerative colitis); no confidence intervals reported	702 records from 372 patients (192 Crohn’s disease, 180 ulcerative colitis)	Indirect	No IBS group; not a test of IBS discrimination
Bennet et al[41], 2018	Multivariate model; predict low-FODMAP responsiveness	61 IBS patients analysed; faecal bacterial profiles	Responder defined as a reduction in IBS-SSS of at least 50 points	None	Responders discriminated from non-responders; no sensitivity, specificity or AUC reported	61 of 67 randomised patients analysed; responder counts not reported by arm	Direct	No discrimination metrics; no external validation
Chumpitazi et al[42], 2015	Association analysis; baseline taxa and low-FODMAP response	33 children with IBS; 48-hour crossover diet periods	Responder defined as at least 50% reduction in abdominal pain frequency	None	Responders enriched at baseline for saccharolytic taxa; no discrimination metric reported	33 children; 48-hour diet periods with 5-day washout	Direct (paediatric)	No predictive model; direction of effect discordant with Bennet et al[41]
Su et al[30], 2026	Deep learning; donor-recipient microbiome matching for FMT	515 FMTs, 30 cohorts, 12 indications including IBS	Clinical response following faecal microbiota transplantation	Retrospective simulation across cohorts	AUC 0.88; 78.7% outcome accuracy; simulated response 49%→71%; no confidence intervals reported	515 FMTs across 30 cohorts and 12 indications	Indirect	Not IBS-specific; retrospective and simulated, not prospective
Dowrick et al[14], 2024	Unsupervised machine learning; subtyping disorders of gut-brain interaction	Population symptom data	Not applicable; evaluates stability of unsupervised subtypes	Not applicable	Demonstrates instability of data-driven subtyping	Not applicable	Direct (DGBI)	Cautionary; subgroups sensitive to methodological choice
Bhagavathula et al[18], 2025	Systematic review; diagnostic accuracy of AI in IBS	6 studies; 1366 participants	Diagnostic accuracy of AI models against the reference IBS diagnosis	Not applicable (systematic review)	Accuracy 54%-98%; AUC 0.61-0.99 across studies; meta-analysis not feasible	6 studies; 1366 participants	Direct (review)	Few studies; variable biomarkers and validation

AI: Artificial intelligence; IBS: Irritable bowel syndrome; OTU: Operational taxonomic unit; AUC: Area under the curve; IBS-C: Irritable bowel syndrome with predominant constipation; IBS-D: Irritable bowel syndrome with predominant diarrhea; FMT: Faecal microbiota transplantation; DGBI: Disorders of gut-brain interaction; FODMAP: Fermentable oligosaccharides, disaccharides, monosaccharides and polyols; IBS^P: Irritable Bowel Syndrome pathogenic-like; IBS^H: Irritable Bowel Syndrome health-like; IBS-SSS: Irritable bowel syndrome severity scoring system.

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	Diagnostic classification; treatment-response prediction	The most common approach; small single-centre cohorts, little external validation
Unsupervised learning (clustering, latent class and mixture models)	Identifies structure without predefined labels	Subtype and endotype discovery	Applied to microbiota and symptom subgrouping; subgroup stability sensitive to method
Deep learning - convolutional neural networks	Hierarchical feature learning from images	Endoscopic image analysis; optical histology prediction	Established in luminal endoscopy; minimal role in functional disease
Deep learning - recurrent networks and transformers	Models sequential or multi-modal data	Longitudinal risk prediction from clinical records	Demonstrated in general clinical records; no published IBS application
Natural language processing and large language models	Extracts structured information from free text	Symptom and outcome extraction from notes and narratives	Demonstrated in inflammatory bowel disease; not yet reported in IBS
Reinforcement learning	Optimises sequential decisions from outcome feedback	Adaptive treatment protocols	Conceptual only in IBS; demonstrated in critical care

IBS: Irritable bowel syndrome.

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	Independent multi-region cohorts; temporal validation; TRIPOD+AI-compliant reporting
Privacy, consent & bias	Sensitive microbiome and psychological data; non-representative data can entrench disparities	GDPR/HIPAA-aligned governance; federated learning; fairness audits on representative datasets
Cost-effectiveness	Development and maintenance are costly; value must be demonstrated	Health-economic modeling; evidence of reduced diagnostic delay and unnecessary colonoscopy
Regulatory landscape	Adaptive and continuously learning tools face evolving oversight	FDA software-as-a-medical-device pathways; predetermined change-control plans; harmonized standards
Patient acceptance and trust	Concerns about bias, transparency, and loss of human connection shape adoption	Transparency, clinician-in-the-loop design, patient education; dedicated acceptance studies

IBS: Irritable bowel syndrome; SHAP: SHapley Additive exPlanations; AI: Artificial intelligence; FDA: Food and Drug Administration; GDPR: General Data Protection Regulation; HIPAA: Health Insurance Portability and Accountability Act; TRIPOD: Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis.

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

IBS: Irritable bowel syndrome; AI: Artificial intelligence; TRIPOD: Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis; PROBAST: Prediction model risk of bias assessment tool; DECIDE: Developmental and exploratory clinical investigations of decision support systems driven by artificial intelligence.