Artificial intelligence in functional gastrointestinal disorders: From precision diagnosis to preventive healthcare

Table 1 Comparison of artificial intelligence applications in irritable bowel syndrome, functional dyspepsia, and gastroesophageal reflux disease

Aspect	IBS	FD	GERD
Common algorithm types	Random Forests and SVM for symptom/microbiome classification; CNN and AutoML for bowel sound and imaging analysis; unsupervised clustering for subtype delineation	Random Forests for TCM pattern identification; unsupervised clustering for symptom subtyping; neural networks for brain-gut imaging; CNN for endoscopic imaging; Ridge Regression for motility prediction	DL (e.g., CNN, ResNet-50) for endoscopic imaging and pH-impedance analysis; ML classification for symptom subtyping and treatment prediction
Primary input data	Symptom questionnaires (Rome IV), psychological scales, EHRs, bowel sound audio, microbiome data, endoscopic images, fNIRS, VOCs	Symptom questionnaires (epigastric pain, postprandial fullness), psychological scales, fNIRS, gastric motility data (GES, manometry), endoscopic images (duodenal), skin conductance, pulse wave	24-hour pH-impedance monitoring, symptom questionnaires, endoscopic images, esophageal manometry, histopathology, DNA methylation biomarkers
Model Objectives	Differentiate IBS from controls (87% accuracy); identify subtypes and novel clusters; predict dietary/drug responses	Differentiate FD from controls (77% accuracy); refine PDS/EPS or TCM patterns; predict prokinetic response (AUC 0.83); identify triggers	Differentiate GERD from functional heartburn (AUC 0.87); detect BE (90% sensitivity); predict PPI/surgical efficacy; assess LES (93.4% accuracy)
Model generalizability	Symptom-psychological clustering applicable to other FGIDs; microbiome and bowel sound models are more specific	Symptom-psychological clustering and brain-gut analysis applicable to IBS; gastric motility models relevant to gastroparesis; TCM models more specific	BE detection models highly specific; pH-impedance and symptom subtyping applicable to other reflux disorders; LES models relevant to motility disorders
Clinical translation progress	Bowel sound diagnosis (87% accuracy)[37]; microbiome models validated[42,43]; digital therapeutics (e.g., Heali App) in trials[44,45]	Early-stage: FNIRS diagnosis (77% accuracy)[50]; endoscopic imaging models (AUC 0.85)[51]; digital therapeutics (e.g., diet apps, AR breathing) in validation[54,55]	More mature: CADe endoscopic systems (90% sensitivity) in trials[60]; pH-impedance analysis (AUC 0.87) near commercialization[57,59]; surgical prediction models validated[66]

AI: Artificial intelligence; IBS: Irritable bowel syndrome; FD: Functional dyspepsia; GERD: Gastroesophageal reflux disease; SVM: Support vector machine; CNN: Convolutional neural network; AutoML: Automated machine learning; TCM: Traditional Chinese medicine; DL: Deep learning; ML: Machine learning; EHRs: Electronic health records; fNIRS: Functional near-infrared spectroscopy; VOCs: Volatile organic compounds; GES: Gastric emptying scintigraphy; GERQ: Gastroesophageal reflux questionnaire; NBI: Narrow-band imaging; PDS: Postprandial distress syndrome; EPS: Epigastric pain syndrome; AUC: Area under the curve; BE: Barrett’s esophagus; PPI: Proton pump inhibitor; LES: Lower esophageal sphincter; FGIDs: Functional gastrointestinal disorders; AR: Augmented reality; CADe: Computer-aided detection.