Artificial intelligence in functional gastrointestinal disorders: From symptom recognition to personalised management

Abstract

Functional gastrointestinal disorders, now termed disorders of gut-brain interaction, are heterogeneous conditions shaped by psychological, physiological, microbial, and environmental factors. Diagnosis and treatment remain difficult because symptoms overlap, trajectories fluctuate, and no single objective biomarker exists. This minireview synthesises human evidence for artificial intelligence (AI) in symptom classification, digital phenotyping, microbiome and multi-omics prediction, treatment-response estimation, wearable monitoring, and language-based care. Targeted PubMed searches were updated through September 20, 2026. Current studies show feasibility, but most use small, single-centre cohorts, retrospective designs, or internal validation; calibration and prospective impact evaluation are uncommon. The most plausible near-term role of AI is therefore clinical augmentation: Identifying longitudinal patterns, supporting shared decisions, and detecting changes that require reassessment while preserving human oversight and patient autonomy. Large language models may improve education and communication, but disorder-specific evaluations reveal imperfect diagnosis and treatment recommendations with marked overconfidence. Translation requires representative cohorts that include overlapping organic disease, explicit endpoints despite the absence of a diagnostic gold standard, external validation, calibration, equity assessment, and prospective evaluation of patient benefit. AI should be judged by outcomes and care experience, not discrimination alone.

Key Words: Artificial intelligence; Machine learning; Functional gastrointestinal disorders; Disorders of gut-brain interaction; Irritable bowel syndrome; Functional dyspepsia; Microbiome; Digital phenotyping; Large language models

Core Tip: This minireview examines the clinical and translational use of artificial intelligence in adults and children with disorders of gut-brain interaction, especially irritable bowel syndrome and functional dyspepsia. It covers machine-learning phenotyping, microbiome and multi-omics prediction, treatment-response estimation, digital monitoring, wearable signals, natural language processing, and large language models. Evidence is interpreted according to study design, sample size, algorithm, validation, calibration, and clinical endpoint; exploratory performance is not treated as proof of clinical benefit.

INTRODUCTION

Functional gastrointestinal disorders, now termed disorders of gut-brain interaction (DGBI), are common, debilitating, and heterogeneous. Rome IV emphasises interactions among motility, visceral sensation, mucosal and immune function, the microbiota, central processing, and psychosocial context rather than a single structural lesion[1,2]. Irritable bowel syndrome (IBS) and functional dyspepsia (FD) are common DGBI; in practice, they overlap, fluctuate over time, and often coexist[3,4]. Population studies demonstrate a substantial global burden on quality of life, work, and health care[5,6].

Biology and life experience provide a specific problem for artificial intelligence (AI). Most of the clinical datasets are organised by disease labels; a patient presenting with abdominal pain, bloating, early satiety, altered stool characteristics, sleep disturbance and anxiety may not fit any single stable label. Thus, for most patients without a disease-defining biomarker, symptom-based diagnosis must still be employed. Therefore, the current guidelines for FD and IBS combine positive symptom criteria with selective investigation, exclusion of alarm features and iterative, preference-sensitive management[7-10]. AI will be useful only if it aligns with this clinical logic. The relevant question is not whether a model can reproduce a Rome category, but whether it can transform high-dimensional, longitudinal information into a more useful account of what is happening for an individual patient and which intervention is most likely to help. Recent discussions in clinical quality and medical education likewise argue that AI should be assessed by how it supports reasoning within a defined workflow and prompts reflective judgement, rather than by accuracy or generic warnings alone[11,12].

This minireview presents human evidence on AI-assisted phenotyping and personalised management for DGBI. We organise the evidence along an end-to-end clinical pathway: Symptom recognition, multidimensional phenotyping, microbiome and multi-omics prediction, treatment-response estimation, language-based support, and patient-centred monitoring (Figure 1 and Table 1). We distinguish proof-of-concept performance from clinical validation and focus on decision support rather than autonomous diagnosis. The structured narrative search and evidence-appraisal approach is described below.

Figure 1 Artificial intelligence-enabled pathway from symptom recognition to personalised management in functional gastrointestinal disorders. Functional gastrointestinal disorders, now termed disorders of gut-brain interaction, include irritable bowel syndrome, functional dyspepsia, and related heterogeneous conditions shaped by physiological, microbial, psychological, and contextual factors. AI: Artificial intelligence.

Table 1 Representative artificial intelligence studies in disorders of gut-brain interaction.

Ref. and task	Data and sample	Algorithm	Reported performance	Validation	Main limitations
Andriulli et al[20], 2003: Dyspepsia classification	Multicentre observational cohort; n = 860	Artificial neural networks	Sensitivity 83.4% for functional dyspepsia; specificity 87.6% for organic disease	Internal dataset evaluation	Pre-Rome IV labels; no calibration, external validation, or impact study
Mousavi et al[18], 2023: DGBI phenotyping	Population-based cross-sectional cohort; n = 3765	Mixed-data spectral clustering	Eight overlapping clusters after re-clustering	Internal cluster indices	Rome III; cross-sectional; no treatment or external validation
Katsumata et al[21], 2023: Diagnostic support	Brain activity and food preference; n = 41, including 25 DGBI	Artificial neural network	Accuracy 72.3% for DGBI and 77.1% for functional dyspepsia	Internal only	Very small sample; control comparison; no calibration
Zhang et al[27], 2025: IBS classification	Resting-state functional magnetic resonance imaging; 31 IBS and 30 controls	Support vector machine	Accuracy 91.9%	Internal validation	Single centre; small sample; no external calibration
Fukui et al[36], 2020: Microbiome classification	16S sequencing; 85 IBS and 26 controls	Feature selection plus machine learning	Sensitivity > 80%; specificity > 90%	Internal validation	Small, region-specific cohort; no independent validation
Yin et al[42], 2022: Acupuncture response	Two clinical trials; total n = 745; external cohort n = 193	Support vector classification and regression	Internal accuracy 0.773; R2: 0.446 for quality of life and 0.413 for symptoms	Independent cohort from a second trial	Predicts prognosis under one treatment, not comparative treatment benefit
Jacobs et al[41], 2021: Cognitive behavioural therapy response	Imaging n = 84; microbiome n = 34	Random forest with 11 genera	Internal area under the receiver operating characteristic curve: 0.96	Internal only	Very small microbiome sample; high optimism risk; no calibration
Tunali et al[53], 2024: Personalised diet	Multicentre randomised trial; n = 121	Microbiome-based diet algorithm	IBS severity change -112.7 vs -99.9; between-group P = 0.29	Prospective randomised comparison	No superiority for the primary outcome; algorithm transparency limited
Ahn and Parker[45], 2026: DGBI language-model benchmark	68 Rome IV clinical scenarios; five models	General-purpose large language models	Diagnostic accuracy 65%-74%; treatment accuracy 41%-54%; confidence 93%-96%	Scenario benchmark only	Convenience sample; no patient outcomes; marked overconfidence

DGBI: Disorders of gut-brain interaction; IBS: Irritable bowel syndrome.

LITERATURE SEARCH AND EVIDENCE APPRAISAL

Targeted PubMed searches were conducted from database inception to September 20, 2026. Search combinations paired DGBI, functional gastrointestinal disorders, IBS, or FD with AI, machine learning, natural language processing, large language model, digital phenotyping, wearable, microbiome, or multi-omics. We prioritised human primary studies, systematic reviews, guidelines, and consensus statements that reported a clinically relevant algorithm, validation strategy, or patient outcome. Animal-only reports, non-gastrointestinal applications, and purely technical studies without a DGBI-relevant outcome were excluded. Reference lists of included sources were hand-searched for additional eligible reports.

Because the included studies were heterogeneous in task, modality, outcome, and validation, we performed a structured narrative synthesis rather than a systematic review or quantitative meta-analysis. Prediction studies were appraised against the main PROBAST domains: Participants, predictors, outcome, and analysis[13]. Table 1 records study design, data source, sample size, algorithm, performance, validation, and study-specific limitations. Table 1 therefore presents a purposive, not exhaustive, selection of nine studies spanning the major clinical pathways and contrasting internal-only validation with independent or prospective validation. These judgements are narrative domain-level appraisals and not formal item-level PROBAST ratings. No claim of exhaustive coverage is made, and a PRISMA flow diagram was therefore not used.

WHY DGBI REQUIRE MULTIDIMENSIONAL MODELS

A useful DGBI model should represent at least three forms of heterogeneity. First, phenotypic heterogeneity reflects differences in symptom expression and severity. FD includes postprandial distress and epigastric pain syndromes with partly distinct physiological correlates, while IBS subtypes based on bowel habit do not capture the full clinical spectrum[3,4,7,8]. Second, mechanistic heterogeneity includes altered motility, visceral hypersensitivity, impaired gut-brain signalling, immune activation, microbial metabolism, and food-related responses. Third, contextual heterogeneity includes stress, mood, sleep, trauma, health beliefs, dietary restriction, social environment, and access to care. Overlap between IBS and FD is associated with greater disease severity and poorer quality of life, so isolated disease labels may miss clinically relevant subgroups[14-16].

Machine learning is suitable for this problem because it can incorporate various types of difficult-to-integrate variables, such as free-text histories, repeated symptom scores, laboratory data, medication exposure, wearable signals, neuroimaging and microbiome features[17-19]. However, complexity is not the same as clinical utility. A model can find a statistical cluster without a reason, and a high-discriminatory model may only encode referral patterns, clinician preferences, or unequal access to tests. Therefore, the design target should be a patient-important decision, such as identifying a need for further examination, supporting a positive diagnosis after reasonable examination, choosing between gut-directed and brain-directed therapy, or detecting a loss of control requiring re-evaluation.

SYMPTOM RECOGNITION AND DIGITAL PHENOTYPING

Early AI studies of dyspepsia illustrate both promise and limited translation. In 2003, neural-network models were evaluated in 860 patients with uninvestigated dyspepsia; the best models reported 83.4% sensitivity for FD and 87.6% specificity for organic disease[20]. However, the labels predated Rome IV, validation was internal to the study dataset, calibration and clinical utility were not reported, and no prospective workflow evaluation followed. These limitations help explain why the approach did not become routine care. More recent studies combined brain activity with food-preference data in 41 participants, including 25 with DGBI, and achieved 72.3% accuracy for DGBI and 77.1% for FD using an artificial neural network[21]. Other studies used supervised or unsupervised learning to identify symptom features and traditional diagnostic patterns[22,23]. These models remain hypothesis-generating and should not replace endoscopy, laboratory evaluation, or clinical judgement when alarm features are present.

For IBS and other DGBI, unsupervised methods may be more appropriate than a single binary classifier. A cross-sectional analysis of 3765 individuals with at least one Rome III disorder used a mixed-data spectral clustering method to identify established and overlapping symptom groups, while a related analysis clustered 988 participants with IBS into nine groups using gastrointestinal and psychological features[18,19]. A prospective phenogroup study linked symptoms and breath-gas profiles in 300 patients with confirmed DGBI, but the supervised gas-production model was internally validated only[24]. Deep phenotyping of patient-generated narratives may capture unpredictability, social restriction, and illness meaning that standard scales omit[25]. These clusters require independent replication and evidence that cluster-guided care improves outcomes.

Digital phenotyping can convert the fluctuating course of DGBI into a longitudinal signal. Smartphone diaries, ecological momentary assessment, food logs, sleep, physical activity and medication records may reveal temporal relationships between stressors, meals, bowel habits and pain. The clinical value is not the volume of data but the ability to identify a reproducible pattern that changes management. Digital therapeutics for IBS have shown that organised self-management can improve symptoms and quality of life, but the intervention itself does not need to be AI-driven[26]. AI can be used to personalise the time and content of prompts, identify non-responses, and recommend when a clinician should review the trajectory. The system should have a small logging cost and return an easy-to-understand summary for the patient, not use continuous self-monitoring as the goal.

Brain-based models offer another form of phenotyping. A single-centre study of 31 patients with IBS and 30 controls used support vector machines with functional-connectivity features and reported 91.9% classification accuracy under internal validation[27]. A smaller study combining prefrontal activity and food preference reported 77.1% accuracy for FD[21]. These findings are biologically plausible because DGBI involve interoception, salience, emotion regulation, and sensorimotor processing. Nevertheless, small samples, flexible imaging pipelines, centre-specific acquisition, and the absence of external calibration make the reported accuracies unstable. Neuroimaging models should therefore be treated as research tools until they are replicated across scanners, populations, and clinical workflows.

MICROBIOME AND MULTI-OMICS PREDICTION

The gut microbiome is a good substrate for AI because it is high-dimensional data that may contain information on host-microbe interactions related to motility, fermentation, immune signalling and visceral sensation. These observations fit within the broader gut-brain-microbiota framework while also highlighting that microbiome results are biologically plausible but clinically varied[28-30]. Studies have associated microbial signatures with IBS severity and have reported compositional differences across IBS subtypes, but systematic reviews also show substantial inconsistency caused by diet, geography, medication, sequencing platform, sample handling and analytic choices[31-34]. Thus, a microbial feature should be regarded as a context-dependent predictor rather than an all-purpose IBS biomarker.

Machine learning can support discovery or prediction. In discovery, models identify microbial associations that separate patients from controls or distinguish symptom groups. A pooled re-analysis of 439 stool samples from six 16S ribosomal RNA datasets used AdaBoost and random forest models with ten-fold cross-validation to identify reproducible genus-level signals, although performance varied across source datasets[35]. In a separate Japanese cohort of 85 patients with IBS and 26 controls, least absolute shrinkage and selection operator feature selection followed by machine learning achieved sensitivity above 80% and specificity above 90%, but validation remained internal[36]. Multi-omics studies have integrated taxonomic, functional, and metabolomic data across bowel-habit subtypes[37]. Paediatric studies further suggest that microbial composition may relate to IBS phenotype and response to fermentable-carbohydrate restriction[38-40].

The clinical appeal is personalised management. Microbial biomarkers in children with IBS were associated with response to a low-fermentable oligosaccharides, disaccharides, monosaccharides, and polyols diet, but the trial included only 33 completers and did not validate a deployable prediction rule[39,40]. In a multimodal cognitive behavioural therapy cohort, 84 patients underwent imaging and only 34 provided microbiome data. A random forest using 11 microbial genera reported an internal area under the receiver operating characteristic curve of 0.96 for treatment response[41]. The model had no independent external or prospective validation, so the estimate is vulnerable to optimism. These studies support biological plausibility and treatment heterogeneity, not routine microbiome testing.

The microbiome is also a typical AI problem: The prediction is correct but for the wrong reasons. A model may have learned a laboratory batch, dietary pattern or antibiotic exposure, but not a treatment-relevant microbial pathway. DGBI research should therefore pre-specify sampling and preprocessing, measure key confounders, report missingness, use nested validation and test transportability across populations. Functional readouts, metabolomics and host data may be more robust than taxonomic abundance alone, but adding modalities increases cost and the risk of overfitting. The goal is a parsimonious, clinically actionable model with calibrated uncertainty.

TREATMENT-RESPONSE ESTIMATION AND PERSONALISED MANAGEMENT

Personalisation requires moving from label prediction to a counterfactual question: What would happen if this patient received treatment A rather than treatment B? Most DGBI AI studies have not reached this standard. A support vector machine study used routine clinical features from 745 patients with FD enrolled in two acupuncture trials. In the internal test set, response classification accuracy was 0.773, while explained variance was 0.446 for quality-of-life improvement and 0.413 for symptom improvement; 193 patients from the second trial formed an independent validation cohort[42]. Even with external validation, the model predicts prognosis under acupuncture rather than comparative benefit against another treatment. A future impact study must compare model-guided selection with guideline-based care and measure whether using the prediction improves outcomes.

Behavioural therapy is a typical case of mechanistic personalisation. Cognitive behavioural therapy can reduce the severity of treatment-resistant IBS, and internet-delivered cognitive behavioural therapy has shown long-term advantages in a systematic review and meta-analysis[43,44]. In a multimodal cohort, baseline microbial features predicted cognitive behavioural therapy response, and responders showed coordinated changes in brain connectivity and microbial composition[41]. AI could eventually support a stepped-care pathway by identifying patients who may benefit from early brain-gut therapy, flagging barriers to engagement, and adapting digital content. It should not be used to infer that psychological factors make symptoms less real; rather, psychological, neural and gastrointestinal signals should be treated as interacting components of the disorder.

Personalised management also includes monitoring. A patient-oriented system could summarise symptom trajectory, stool pattern, sleep, diet, medication exposure and quality-of-life change, then identify clinically meaningful deterioration or non-response. The output should be a small number of actionable signals, such as a sustained deviation from the baseline, a potential adverse effect, a new alarm feature or a need to re-examine the diagnosis. Patient-reported outcomes are still the focus, as symptom improvement is the desired effect. The system should allow patients to correct the record, select what is to be shared, and opt out of passive data collection.

A safe-care path would have AI assisting in a supportive capacity. The model will propose some possible phenotypes or treatments, but a doctor will consider red flags, other illnesses, medications, pregnancy status, nutritional risks and the patient’s wishes. Shared decision-making should be used for cases of dietary restriction, neuromodulation and psychological therapy. Models should show calibrated probabilities and the features that substantially affected the output, not a single opaque recommendation.

LARGE LANGUAGE MODELS AND LANGUAGE-BASED CARE

Large language models broaden the AI opportunity because functional gastrointestinal disorder care is rich in language. Patients describe pain, urgency, food fear, stool changes and illness-related worry in narratives that are difficult to reduce to checkboxes. An large language models could structure a history, identify missing information, summarise longitudinal notes, translate patient education, generate a draft symptom diary or facilitate asynchronous communication. These applications are plausible extensions of natural language processing, but they should not be autonomous diagnostic or treatment-prescribing systems.

Evidence is now emerging in DGBI rather than only in general medicine. In a 2026 benchmark, five general-purpose large language models were tested on 68 Rome IV multidimensional clinical scenarios. Diagnostic accuracy ranged from 65% to 74%, treatment accuracy from 41% to 54%, and stated confidence remained 93% to 96% regardless of correctness[45]. This mismatch between performance and confidence is particularly concerning in DGBI, where symptoms overlap with organic disease and clinical context determines whether reassurance is safe. Reviews of gastroenterology applications also report heterogeneous tasks, inconsistent evaluation, and limited prospective evidence[46-48]. Large language models may help structure symptom chronology and patient goals, but unverified outputs should not determine diagnosis or treatment.

A clinically responsible architecture would ground responses in an approved knowledge base, display supporting passages, communicate uncertainty, and escalate alarm features or ambiguous cases to a clinician. Hallucination is not merely a generic technical defect in DGBI. A model could falsely reassure a patient that symptoms are "just stress", invent restrictive low-FODMAP rules, overlook nutritional risk, or delay investigation of bleeding, weight loss, anaemia, or progressive dysphagia[49]. In an IBS dietary-advice evaluation, 70% of ChatGPT responses and 57.5% of Gemini responses were guideline-compliant, leaving substantial room for unsafe or incomplete advice[50]. Prompt and output logs should be governed as clinical data, and patient-facing systems require prospective testing across language, literacy, disability, culture, and access.

ADJACENT AND EMERGING APPLICATIONS

Natural language processing could support electronic health record case finding by extracting Rome symptoms, alarm features, prior investigations, and longitudinal trajectories. However, our search identified no prospectively validated, DGBI-specific electronic health record case-finding system. Models trained on diagnostic codes would risk reproducing coding behaviour and circular labels. Development should therefore use clinician-adjudicated reference standards, explicitly represent overlapping organic disease, and validate transportability in an independent health system[48].

AI has progressed further in adjacent functional testing. A multi-stage model trained on 1741 high-resolution oesophageal manometry studies used convolutional neural networks and feature-based models, achieving top-1 diagnostic accuracy of 0.81 and top-2 accuracy of 0.92 on a held-out test set[51]. This demonstrates that physiological traces can be standardised algorithmically, but oesophageal motility disorders are not interchangeable with DGBI, and comparison with expert labels does not establish improved patient outcomes. Similar caution applies to proposed gastric-emptying or other motility applications.

Wearable monitoring remains exploratory. A 2026 systematic review included 36 studies and 3986 patients with DGBI. Heart-rate-variability measures differed modestly from healthy controls, but associations with symptom severity were heterogeneous and no eligible skin-conductance study was identified[52]. Thus, electrocardiography, accelerometry, sleep, and electrodermal signals should be treated as candidate longitudinal features rather than validated flare detectors. Future work needs patient-level temporal validation and predefined actions triggered by a detected change.

Digital therapeutics and dietary support illustrate the same distinction between digital delivery and adaptive AI. Zemedy improved IBS self-management but was not an AI-adaptive intervention[26]. In a multicentre trial of 121 patients, a microbiome-based AI-assisted personalised diet improved symptoms but did not outperform a standard low-FODMAP diet for the primary between-group outcome[53]. Generative dietary advice is even less mature and should remain dietitian-supervised because guideline compliance is incomplete[50].

BARRIERS TO TRANSLATION

Data quality is the first barrier. Functional gastrointestinal disorder labels are partially clinical constructs, symptom reporting is affected by language and expectation, and missingness is informative because patients with more complex disease often have more encounters. Retrospective electronic health records therefore contain a combination of biology, clinician behaviour and healthcare access. Future cohorts should use the harmonised Rome criteria, standardised patient-reported outcomes, explicitly collect alarm features and conduct longitudinal follow-up. Data linkage should be transparent about what is directly measured and what is inferred.

Generalisation is the second barrier. External validation is uncommon in prediction research, and performance often declines at a new hospital or in a different population[54,55]. DGBI models are particularly sensitive to diet, culture, referral pathway, language, socioeconomic status, and local testing or prescribing practices. Applying PROBAST domains to the representative studies in Table 1 identifies recurring risks: Small samples relative to predictor count, single-centre recruitment, internally selected features, incomplete handling of missing data, weak or circular reference standards, and reliance on discrimination without calibration or decision analysis[13]. Research should report calibration, subgroup performance, and net benefit in addition to area under the receiver operating characteristic curve. High accuracy against retrospective labels does not guarantee clinical benefit[56,57].

Interpretability is the third problem. Explanation methods can show how a model works in a more understandable way; however, they are generally not causal mechanisms and may vary from case to case[58,59]. Functional gastrointestinal disorders should not be taken as proof that a food, mood, or microbe caused the symptoms based on feature attribution. A convenient way to explain it is operationally: What information was unavailable, what pattern was recognised, how uncertain was the prediction, and what alternative actions were considered? Prospective studies will determine whether these explanations assist clinicians in their work and inform patients.

Equity, privacy and accountability are inseparable from implementation. AI will widen the gap if high-quality data are mainly from tertiary centres, and if digital monitoring requires a smartphone, stable internet and high health literacy. Health-equity safeguards should be built into cohort recruitment, model development and performance reporting[60]. Gastroenterology position statements and ethical analyses emphasise transparency, informed consent, data governance, clinician responsibility and the need to keep the patient at the centre[61-64]. The clinician should remain accountable for care, while the institution must own monitoring, incident reporting and model updates.

Deployment also needs proof after model development. General AI frameworks in medicine and gastroenterology commentary stress that useful systems augment clinical work rather than solely automating predictions[65-68]. A pragmatic functional gastrointestinal disorder trial could randomly assign clinics or patients to guideline-based care with or without AI-supported phenotyping to measure symptom response, quality of life, diagnostic safety, clinician workload, equity and cost. Post-deployment monitoring will track calibration drift, subgroup failures, alert load and patient-reported trust. A model that improves area under the receiver operating characteristic curve but increases anxiety, unnecessary restriction or clinician workload is not a successful intervention. Accordingly, clinical deployment should evaluate the process of AI-supported reasoning and its fit within a defined workflow, while educational use may be most valuable when AI disagreement is used to prompt reflective rather than deferential decision-making[11,12].

A TRANSLATIONAL ROADMAP

The next generation of DGBI AI studies should follow a disorder-specific sequence. First, define the clinical decision and an endpoint that does not assume a perfect diagnostic gold standard. A positive-diagnosis endpoint should combine prespecified symptom criteria, appropriate exclusion of organic disease, longitudinal diagnostic stability, and safety outcomes such as missed alarm pathology. Second, recruit representative longitudinal cohorts that include DGBI overlap, common organic mimics, alarm features, and patients whose diagnosis changes over time. Third, lock preprocessing, missing-data handling, and feature availability at the intended decision point. Fourth, perform leakage-controlled internal validation followed by external validation across institutions, languages, and care settings. Fifth, report calibration, net benefit, and subgroup equity, not discrimination alone. Sixth, conduct a prospective impact study that allows clinicians and patients to accept or reject model advice. Seventh, monitor diagnostic delay, dietary harm, anxiety, workload, calibration drift, and subgroup failures after deployment.

Multimodal integration should be used only when it adds measurable value. A longitudinal symptom trajectory may transport better than a microbiome-plus-imaging model, and a transparent model may be preferable for a low-risk, reversible decision. Federated or privacy-preserving learning may support multi-centre collaboration, but it does not solve inconsistent labels or biased recruitment. Large language models should begin with restricted tasks such as summarisation, question generation, and retrieval-grounded education. Higher-risk diagnostic or treatment functions require human review and prospective safety evaluation.

Shared benchmarks should reflect the decisions that matter in DGBI. Candidate outcomes include time to a safe positive diagnosis, missed organic disease, unnecessary testing, symptom severity, adequate relief, quality of life, nutritional safety, adherence, access, decisional conflict, clinician workload, and cost. For clinician-facing systems, the intervention is the clinician-patient-model team rather than the algorithm alone. Trials should therefore compare care teams with and without AI support and measure whether recommendations change, whether patients understand and agree with them, and whether those changes improve outcomes. Benchmark datasets should include overlapping DGBI, relevant organic disease, and alarm features so that a safe model learns when not to reassure.

CONCLUSION

AI offers a credible route to more multidimensional, longitudinal, and personalised DGBI care. Current evidence supports three cautious conclusions. First, machine learning can identify symptom and mechanistic subgroups that conventional labels may not capture. Second, microbiome, brain, behavioural, and longitudinal physiological data may help explain variation in response to diet, acupuncture, cognitive behavioural therapy, and other interventions. Third, large language models may improve the organisation and accessibility of language-rich care, but DGBI-specific evaluations show incomplete diagnostic and treatment accuracy with marked overconfidence.

The central limitation is the distance between exploratory prediction and demonstrated patient benefit. Small or homogeneous datasets, unstable labels, internal validation, unreported calibration, and uncertain causal interpretation limit translation. The most defensible near-term role is AI as a transparent co-pilot that helps clinicians and patients recognise patterns, quantify uncertainty, and follow change. Clinicians must retain responsibility for alarm-feature assessment, organic-disease exclusion, nutritional safety, and shared decisions. Success should be measured by safer diagnosis, better outcomes, and improved care experience, not technical performance in isolation.