Artificial intelligence in gastroenterology practice: Current perspectives and emerging prospects

Abstract

Artificial intelligence (AI) is increasingly changing the landscape of gastroenterology by improving disease detection, diagnosis, risk stratification, prognostication, and clinical decision-making. Machine learning, deep learning, natural language processing, large language models, and multimodal data integration have facilitated the development of accurate diagnostic and predictive algorithms for a wide spectrum of gastrointestinal and hepatopancreatobiliary diseases. In upper gastrointestinal diseases, AI has enhanced the detection and characterization of gastroesophageal reflux disease, Barrett’s esophagus, esophageal cancer, gastric neoplasia, and gastroesophageal varices. The use of AI in capsule endoscopy has facilitated small intestinal mucosal lesion detection and reduced reading time. Computer-aided detection and diagnosis systems have improved adenoma detection and polyp characterization in colorectal diseases, while AI is increasingly being applied to molecular profiling and prognostic modeling in colorectal cancer. In inflammatory bowel disease, AI facilitates more objective endoscopic and histologic assessment, as well as prediction of disease outcome and treatment response. Promising AI-based radiomics, digital pathology, biomarker discovery, and risk prediction models have been emerging in pancreatic, biliary, and liver diseases. Large language models and multimodal AI may further enable integration of heterogeneous information and support personalized decision-making. Computer-aided quality assessment represents an emerging application for standardizing procedural quality in endoscopy. Several AI-assisted platforms have received regulatory approval with ongoing integration into clinical practice. However, limited external validation, population and device-related heterogeneity, dataset bias, interpretability, workflow integration, and uncertain cost-effectiveness remain important barriers to generalizability and clinical adoption. Robust multicenter prospective validation and responsible implementation will be essential to optimize its clinical impact.

Key Words: Artificial intelligence; Machine learning; Deep learning; Gastroenterology; Endoscopy; Capsule endoscopy; Inflammatory bowel disease; Colorectal cancer; Pancreatic diseases; Hepatobiliary diseases

Core Tip: Artificial intelligence (AI) is increasingly reshaping the landscape of gastroenterology by improving disease detection, endoscopic assessment, risk stratification, prognostication, and therapeutic decision-making. This review provides an overview of current AI applications across gastrointestinal, pancreatic, and hepatobiliary diseases, highlights United States Food and Drug Administration-approved AI platforms integrated into clinical practice, and discusses key limitations and future directions toward precision gastroenterology.

INTRODUCTION

Gastroenterology, being a data-rich discipline, has increasingly relied on the interpretation of complex clinical, imaging, histopathological, and molecular information. As a consequence, artificial intelligence (AI) has emerged as a powerful tool for improving disease detection, prognostication, and clinical decision-making across gastrointestinal and hepatobiliary disorders.

AI encompasses computational systems that have human-like capabilities for pattern recognition, prediction, and decision-making, as well as the ability to understand natural language. Machine learning (ML), the foundational subset of AI, learns patterns from structured datasets such as demographic profiles, clinical characteristics, laboratory or imaging data, and molecular information to facilitate disease diagnosis, risk stratification, and clinical decision-making. Common ML models include support vector machines (SVMs) for medical classification tasks and artificial neural networks (ANNs) for modeling complex nonlinear relationships between clinical variables and outcomes[1]. Deep learning (DL) is an advanced subset of ML that employs ANNs with multiple hidden layers to automatically learn hierarchical features from large and complex datasets[1,2]. Major DL architectures include convolutional neural networks (CNNs) for medical image analysis, recurrent neural networks and long short-term memory networks for sequential and temporal data, and transformer models, which leverage self-attention mechanisms and now represent the state of the art in both computer vision and natural language processing (NLP)[2-4]. NLP applies a range of computational algorithms, including ML and DL, to extract structured information from unstructured clinical text, including electronic health records (EHRs), pathology and radiology reports, endoscopy notes, and discharge summaries[1,4]. Large language models (LLMs) represent an advanced application of NLP, leveraging DL and large-scale pretraining to understand, generate, and synthesize complex clinical information from diverse sources[4].

Gastroenterology presents unique challenges for AI because it generates diverse, high-dimensional data, including real-time endoscopic videos, histopathology, radiologic imaging, and clinical text[1,3]. Endoscopic procedures produce thousands of image frames per examination, requiring CNN- and transformer-based models capable of real-time feature extraction despite motion artifacts, variable illumination, and inconsistent bowel preparation[5,6]. Whole-slide histopathology and multiparametric imaging further increase data complexity, while NLP facilitates information extraction from unstructured clinical records, necessitating specialized AI approaches tailored to gastroenterology[1,3,7]. The key AI technologies and their applications in gastroenterology are illustrated in Figure 1.

Figure 1 Overview of key artificial intelligence technologies and their applications in gastroenterology. AI: Artificial intelligence; ML: Machine learning; SVM: Support vector machine; ANN: Artificial neural network; DL: Deep learning; CNN: Convolutional neural network; RNN: Recurrent neural network; LSTM: Long short-term memory; LLM: Large language model; NLP: Natural language processing.

A thorough literature search was conducted to ensure a structured and transparent selection of the evidence. The following electronic databases were searched: PubMed/MEDLINE, Scopus, Web of Science, and the Cochrane Library. The search terms included: “artificial intelligence”, “machine learning”, “deep learning”, “natural language processing”, “large language models”, “computer-aided detection”, “computer-aided diagnosis”, “endoscopy”, “capsule endoscopy”, “gastroesophageal reflux disease”, “Barrett’s esophagus”, “esophageal cancer”, “Helicobacter pylori”, “gastritis”, “gastric cancer”, “celiac disease”, “small bowel obstruction”, “inflammatory bowel disease”, “Crohn’s disease”, “ulcerative colitis”, “colorectal polyp”, “colorectal cancer”, “acute pancreatitis”, “chronic pancreatitis”, “pancreatic cystic neoplasm”, “pancreatic cancer”, “choledocholithiasis”, “biliary stricture”, “post-ERCP pancreatitis”, “liver fibrosis”, “cirrhosis”, “steatotic liver disease”, “viral hepatitis”, “autoimmune hepatitis”, and “hepatocellular carcinoma”. These terms were combined with Boolean operators (AND, OR) to increase the reach. The publication period was restricted to articles published between January 2003 and September 2026. Peer-reviewed original research articles, systematic reviews, meta-analyses, and narrative reviews published in the English language were included. Letters, editorials, conference abstracts without full-text availability, and non-peer-reviewed preprints were excluded. Reference lists of selected articles were hand-searched for additional relevant publications.

AI IN ESOPHAGEAL DISEASES

Gastroesophageal reflux disease

AI applications in gastroesophageal reflux disease (GERD) have evolved from symptom-based predictive models to automated endoscopic assessment systems. Early ANN models accurately distinguished GERD from healthy controls based on symptom profiles but had limited ability to differentiate erosive from non-erosive disease[8]. Subsequent image-based ML and DL approaches, including SVM, CNN-based GERD-VGGNet and GerdNet-RF models, improved automated endoscopic diagnosis, with reported performance comparable to or better than trainee endoscopists[9-11]. However, small cohorts, retrospective designs, overfitting, limited external validation, and equipment-specific validation constrain generalizability. DL systems have also automated grading of the gastroesophageal flap valve using the Hill classification, although performance was lower for adjacent grades and validation remained largely single-center[12]. AI has also shown promise in automated pH-impedance analysis. An AI-based baseline impedance ratio demonstrated good agreement with conventional measurements and improved prediction of treatment response in borderline cases[13]. DL models based on ResNet and Structured State Space sequence (S4) frameworks achieved expert-level proficiency in reflux episode detection and post-reflux swallow-induced peristaltic wave characterization, effectively reducing false-positive detections and the burden of manual interpretation, although the studies were limited by small cohorts and the continued need for manual analysis[14,15].

Barrett’s esophagus

AI has significantly enhanced the non-invasive screening, risk stratification, and histopathologic assessment of Barrett’s esophagus (BE). ML models using clinical risk factors and EHR data demonstrated moderate-to-good discrimination for BE and esophageal adenocarcinoma, with external or holdout area under the receiver operating characteristic curve around 0.81-0.84, although case-control enrichment and limited explainability remain important limitations[16,17]. AI-assisted Cytosponge-Trefoil Factor 3 analysis enables automated identification of clinically significant BE and distinction from focal intestinal metaplasia, potentially reducing unnecessary endoscopy, although small test sets and false-negative cases remain concerns[18]. DL models have also shown promising performance in histological dysplasia classification, although validation has been pilot-scale and real-world community pathology performance remains uncertain[19].

Esophageal cancer

AI has contributed significantly to the risk prediction, detection, characterization, staging, and prognostication of esophageal malignancies. ML models combining cytological and epidemiological data accurately identify esophageal squamous cell carcinoma (ESCC) and esophagogastric junction adenocarcinoma while reducing unnecessary endoscopic procedures, with validation in a nationwide prospective multicohort study involving over 17000 participants[20]. DL-based endoscopic systems showed high sensitivity for BE-associated neoplasia detection and demonstrated good performance for superficial ESCC detection and delineation in multicenter and prospective evaluation[21,22]. However, accurate lesion-margin delineation and the scarcity of subtle neoplastic lesions remain challenges. DL models have also achieved high accuracy for endoscopic classification of BE, ESCC, and esophageal adenocarcinoma, although based on relatively small single-center datasets[23]. ML-based survival prediction in ESCC showed moderate discrimination but remained limited by retrospective, single-region validation[24].

Achalasia cardia

AI has emerged as a valuable tool for the diagnosis, phenotypic characterization, and treatment stratification of achalasia. A CNN-based model accurately classified esophageal motility disorders from high-resolution esophageal manometry images, although validation was limited by a small, single-center dataset and exclusion of complex cases[25]. A multicenter ML model using clinical, manometric, and radiographic features enabled risk stratification after peroral endoscopic myotomy, but showed only modest predictive performance, particularly for post-peroral endoscopic myotomy reflux esophagitis[26]. ML analysis of functional luminal imaging probe panometry also demonstrated promising discrimination of spastic and non-spastic achalasia, although reduced test-set performance and reliance on human-scored variables limit generalizability[27].

Gastroesophageal varices

AI has shown considerable promise in improving the detection, grading, and bleeding risk assessment of gastroesophageal varices (GEV). Multicenter DL systems demonstrated high accuracy for variceal detection, classification, segmentation, and red-color sign detection, including prospective validation[28,29]. Transfer learning models also predicted 12-month variceal bleeding with good performance, although reliance on single endoscopic images may limit generalizability[30]. Multimodal models combining endoscopic and clinical data achieved high discrimination for bleeding risk, but limited sample size and lack of comparison with established criteria remain important limitations[31].

Overall, the evidence is strongest for AI-assisted detection and characterization of esophageal neoplasia and GEV. In esophageal cancer, risk prediction for ESCC and esophagogastric junction adenocarcinoma has been evaluated prospectively in a large nationwide multicenter cohort, while AI-based detection and delineation of superficial ESCC has undergone multicenter external and prospective validation[20,22]. AI systems for GEV have also undergone multicenter and prospective validation, including automated detection, segmentation, and red-color sign assessment[28,29]. Other applications, including automated Barrett’s neoplasia detection and GERD assessment, have shown promising performance but remain less mature because of limited prospective validation or smaller test cohorts[10,11,18,21]. Evidence for clinical workflow improvement is emerging, particularly where AI can support endoscopic detection or reduce the burden of manual image or physiological interpretation, but direct evidence of improved patient outcomes remains limited. Broader real-world implementation across different clinical settings, equipment, and patient populations has yet to be established. Studies evaluating AI applications in esophageal diseases, including their methodologies, performance metrics, and key limitations, are summarized in Table 1.

Table 1 Studies evaluating applications of artificial intelligence in esophageal diseases.

Ref.	Clinical application	Data type	AI method	Sample size	Study design, center design, and validation design	Major performance metric	Principal limitation
Pace et al[8]	GERD vs controls, EE vs NERD discrimination	QUID questionnaire (clinical/symptom data)	ANN, TWIST evolutionary algorithm	577 patients, 94 controls	Prospective, multicenter (60 centers), resampling-based training, testing, and validation	Accuracy: 99.2% for GERD vs controls, 66.5% for EE vs NERD	Poor discrimination between EE and NERD
Huang et al[11]	GERD diagnosis	Endoscopic images of esophagogastric junction	HHDF-SVM	39 patients, 108 controls	Diagnostic accuracy study, single-center, 10-fold cross-validation	Accuracy: 93.2%, sensitivity: 94.9%, performance exceeded inexperienced endoscopists	Small sample size, no automated LA grading
Wang et al[9]	LA classification of GERD	Conventional and NBI endoscopic images	GERD-VGGNet	496 patients, 464 development cohort, 32 external test cohort	Retrospective, single-center, 10-fold cross-validation, external test set	NBI accuracy: 99.2% internally, 87.5% external testing, external test accuracy exceeded young trainee endoscopists	Small external test set, overfitting without augmentation
Yen et al[10]	LA classification of GERD	NBI endoscopic images	GERD-VGGNet features, random forest	671 development images, 32 test images	Retrospective, single-center, 10-fold cross-validation, independent test set	Accuracy: 98.4% internally, 92.5% independent testing	Retrospective design, small independent test set
Ge et al[12]	Hill classification of gastroesophageal flap valve	RGB endoscopic images	ResNet-50, dual attention (EDA + SE)	3256 images, 2589 for training, 334 for validation, 333 for test	Retrospective, single-center, training, validation, and independent test sets	Accuracy: 93.39%, AUC: 0.989	Single-center, reduced performance for adjacent categories (Hill II vs III)
Rogers et al[13]	Automated baseline impedance and reflux assessment	pH-impedance monitoring data	Decision tree analysis	325 individuals, 117 symptomatic off-PPI, 93 symptomatic on-PPI, 115 controls	Exploratory observational cohort, multicenter, comparison with manual reference standards	Event identification accuracy: 88.5%, AUC: 0.766 for treatment-response prediction, calibration: regression agreement with manual MNBI, r2 = 0.79-0.86	Proof-of-concept, excluded very low baseline impedance cases
Zhou et al[14]	Reflux event identification	24-hour pH-impedance monitoring	Structured State Space (S4) model	45 patients, 25 for training, 20 for test	Retrospective, single-center, training, validation, and independent test sets	AUC: 0.87 (95%CI: 0.85-0.89), specificity: 80.8%, threshold calibrated to expert false-positive rate	Small sample size, relatively limited sensitivity
Wong et al[15]	Reflux episodes and PSPW index assessment	24-hour impedance-pH monitoring	Deep residual network (ResNet)	106 patients	Retrospective analysis, single-center, validation against manual Wingate criteria	Accuracy: 87% for reflux episodes, 82% for PSPW index, calibration: manual Wingate criteria used as reference standard	Clinical adjunct, manual analysis remains necessary
Rosenfeld et al[16]	BE detection/risk prediction	Questionnaire (symptoms/risk factors)	Logistic regression with IG and CFS feature selection	1299 (BEST2/training and internal validation), 398 (BOOST/external validation)	Case-control, multicenter, internal split-sample, external independent cohort	Training AUC: 0.87 (95%CI: 0.84-0.90), internal: 0.86 (95%CI: 0.83-0.89), external: 0.81 (95%CI: 0.74-0.84), calibration: Probability output calibrated to 90% sensitivity, log-loss/precision-recall assessed	Case-control enrichment may overestimate performance, requires prospective primary-care validation
Iyer et al[17]	BE and EAC prediction	EHRs	Ensemble transformer-based ML model	8476 BE cases, 1539 EAC cases, 252276 controls	Retrospective cohort, multicenter, 10% model holdout set	AUROC: 0.84 for both BE and EAC	Limited explainability, requires validation outside Mayo Clinic EHR systems
Berman et al[18]	BE segment length prediction	Digitized whole-slide images (Cytosponge-TFF3)	CNN for tile identification, logistic regression	529 (BEST2), 158 (BEST3)	Case-control, RCT evaluation, multicenter, 5-fold cross-validation, independent RCT cohort	Internal accuracy: 0.84 (95%CI: 0.77-0.90), precision: 0.95 (95%CI: 0.87-1.00), RCT precision: 0.91 (95%CI: 0.85-0.97), clinically relevant BE: 55% could avoid gastroscopy, false-negative rate 5.1%	Small independent test set, 5.1% false-negative rate for clinically relevant BE
Faghani et al[19]	Histological dysplasia classification (NDBE, LGD, HGD)	Digitized histology whole-slide images	Ensemble deep learning (YOLO, ResNet101)	542 patients, 70% for training, 20% for validation, 10% for test	Retrospective validation, random 70/20/10 train/validation/test split	LGD sensitivity: 81.3%, specificity: 100%, NDBE/HGD > 90%, F1-score: 0.90-1.0	Pilot-scale validation, real-world community pathology performance unknown
Gao et al[20]	ESCC and esophagogastric junction adenocarcinoma risk prediction	105 cytological, 15 epidemiological features	Automated ML-based prediction tool	17498 eligible participants	Nationwide, multicohort, prospective, multicenter (39 hospitals), internal testing, community-based validation	Internal AUROC: 0.960 (95%CI: 0.937-0.977), prospective AUROC: 0.964 (95%CI: 0.920-0.990), clinically: Unnecessary endoscopies avoided in 90.3% of the testing cohort and 92.8% of community screening participants	Primarily focused on resource optimization for endoscopy in China
Fockens et al[21]	Detection of early Barrett’s neoplasia	White-light endoscopic images and videos	EfficientNet-Lite1 encoder, MobileNetV2 DeepLabV3+ decoder	14046 images from 2506 patients for training	Model development and validation, multicenter (15 international centers), all-comers and enriched benchmarking test sets	Sensitivity: 95% (95%CI: 88%-98%) for images, 97% (95%CI: 90%-99%) for videos in all-comers set	Training relied on high-quality expert images, subtle neoplastic lesions were scarce
Yuan et al[22]	Detection and delineation of superficial ESCC and precancerous lesions	Non-magnified NBI	AI-based detection and delineation system	10047 still images and 140 videos (1112 patients)	Multicenter development and validation, multicenter (4 hospitals), internal/external image test sets and prospective clinical evaluation	Accuracy: 92.4% internal, 89.9% external, 91.4% prospective	Accurate delineation of lesion margins remains a key clinical challenge
Chempak Kumar et al[23]	Classification and staging of ESCC, EAC, and BE	Endoscopic images	Multi-CNN (Xception, MobileNetV2, GoogLeNet, Darknet53), ABC optimization, SVM	1028 images (523 BE, 217 ESCC, 288 EAC)	CAD network development, single-center, 3-fold cross-validation	Overall classification accuracy: 97.76%	Relatively small dataset
Lu et al[24]	Prognostic prediction in ESCC	27 clinical and laboratory blood parameters	Random survival forest	2521 consecutive patients, 1765 for training and 756 for test	Retrospective, single-center, 70% training/30% test split	C-index: 0.746, 5-year AUC: 0.771	Single-region cohort, retrospective design limits generalizability
Popa et al[25]	Classification of esophageal motility disorders	Single wet-swallow raw HREM images	DL (CNN, Inception V3)	157 patients (1570 images), 70% training, 15% testing, 15% validation	Retrospective, single-center, 70% training/15% testing/15% validation	Overall accuracy: 94% (validation dataset)	Small sample size, diagnosis-purity bias from exclusion of complex cases
Takahashi et al[26]	Prediction of persistent symptoms and reflux esophagitis after POEM in achalasia	Demographic data, disease duration, HRM, esophagography	ML clustering and predictive models	1824 treatment-naïve patients including 1778 undergoing POEM	Multicenter cohort, model development for risk stratification	AUC: 0.70 for persistent symptoms, 0.61 for reflux esophagitis	Weak predictive performance for post-POEM reflux esophagitis
Carlson et al[27]	Classification of achalasia subtypes using FLIP panometry	FLIP panometry	Supervised ML (decision trees)	180 adult patients	Retrospective analysis of a prospective cohort, single-center, 140 training/40 test cohort	Spastic accuracy: 90%, non-spastic accuracy: 78%	Marked reduction in test-set accuracy, limited generalizability due to human-scored variables
Wang et al[28]	GEV segmentation and RC sign detection	Endoscopic images and videos	ENDOANGEL-GEV (U-Net++, DCNN, DBSCAN)	1156 training patients, 161 prospective patients	Multicenter retrospective with prospective observational, internal, multicenter video, and prospective validation	EV segmentation mIoU: 0.8087 (95%CI: 0.7968-0.8206), prospective RC detection accuracy: 94.62% (95%CI: 89.11%-97.56%), clinical impact: Correct prophylactic intervention in 12.3% additional patients and avoidance of unnecessary prophylaxis in 52.6% low-risk patients	Single-arm design, potential bias from high proportion of severe Child-Pugh C cases
Chen et al[29]	GEV detection and classification	Endoscopic images and videos	ENDOANGEL (DCNN)	6189 patients with 14718 images	Multicenter retrospective, independent image and multicenter video validation	EV detection accuracy: 97.00%, GV detection accuracy: 92.00%	Focused mainly on detection/classification, lacked later interpretable logic
Hong et al[30]	Prediction of 12-month EV bleeding	Endoscopic images	DL (transfer learning with EfficientNet)	1075 images (675 training/validation, 400 test)	Multicenter retrospective cohort, internal and external validation	EfficientNet accuracy: 0.893, recall: 0.870	Reliance on single endoscopic images rather than multiple views or clinical data
Wang et al[31]	Prediction of 12-month EV bleeding	Endoscopic images and clinical structured data	MMML stacking model (EfficientNet + 6 algorithms)	341 patients, 161 patients in the external test cohort	Multicenter retrospective cohort, internal validation and external testing	AUC: 0.975, sensitivity: 0.952	No comparison with Baveno criteria, no elastography data, limited sample size

AI: Artificial intelligence; GERD: Gastroesophageal reflux disease; EE: Erosive esophagitis; NERD: Nonerosive reflux disease; QUID: Questionario Italiano Diagnostico; ANN: Artificial neural network; TWIST: Training and testing and input selection; HHDF-SVM: Hierarchical heterogeneous descriptor fusion support vector machine; LA: Los Angeles; NBI: Narrow-band imaging; VGGNet: Visual Geometry Group network; RGB: Red, green, blue; EDA: Early dynamic attention; SE: Squeeze-and-excitation; AUC: Area under the curve; PPI: Proton pump inhibitor; MNBI: Mean nocturnal baseline impedance; CI: Confidence interval; PSPW: Postreflux swallow-induced peristaltic wave; BE: Barrett’s esophagus; IG: Information gain; CFS: Correlation-based feature selection; EAC: Esophageal adenocarcinoma; EHR: Electronic health record; ML: Machine learning; AUROC: Area under the receiver operating characteristic curve; TFF3: Trefoil factor 3; CNN: Convolutional neural network; RCT: Randomized controlled trial; NDBE: Nondysplastic Barrett’s esophagus; LGD: Low-grade dysplasia; HGD: High-grade dysplasia; ESCC: Esophageal squamous cell carcinoma; ABC: Artificial bee colony; SVM: Support vector machine; CAD: Computer-aided diagnosis; HREM: High-resolution esophageal manometry; DL: Deep learning; POEM: Peroral endoscopic myotomy; HRM: High-resolution manometry; FLIP: Functional luminal imaging probe; GEV: Gastroesophageal varices; RC: Red color; EV: Esophageal varices; mIoU: Mean intersection over union; GV: Gastric varices; MMML: Multimodal machine learning; BEST2: Barrett’s Esophagus Screening Trial 2; BEST3: Barrett’s Esophagus Trial 3; BOOST: BOOST case-control study; C-index: Concordance index; DBSCAN: Density-based spatial clustering of applications with noise; DCNN: Deep convolutional neural network; ResNet: Residual network; r²: Coefficient of determination; S4: Structured state space sequence model; U-Net++: Nested U-Net; YOLO: You Only Look Once.

AI IN GASTRIC DISEASES

AI has been successfully incorporated into the detection of Helicobacter pylori (H. pylori) infection, precancerous lesions, and gastric neoplasia. Early CNN-based models achieved 87.7% accuracy for detecting H. pylori from endoscopic images, while a meta-analysis reported pooled accuracy, sensitivity, and specificity of 87.1%, 86.3%, and 87.1%, respectively[32,33]. DL approaches for detecting gastric atrophy and intestinal metaplasia from endoscopic images or videos showed moderate-to-high diagnostic performance, although validation was largely single-center and histological confirmation was limited[34,35]. Early CNN-based models established the feasibility of automated gastric cancer detection from endoscopic images, while subsequent large-scale validation studies confirmed diagnostic performance comparable to expert endoscopists and superior to non-experts, with high sensitivity for early gastric cancer detection[36-38].

AI has also shown promise for early gastric cancer localization and prediction of submucosal invasion, with the latter achieving 89.3% accuracy in prospective external validation[39,40]. Meta-analytic evidence supports this excellent performance of AI for both gastric cancer diagnosis and invasion-depth prediction[41]. Recent real-time AI systems enabled automated lesion detection and classification, but improvements in detection rates over conventional endoscopy were not statistically significant[42]. Two single-center randomized clinical trials (RCTs) have demonstrated discernible clinical utility with improved detection of gastric atrophy, intestinal metaplasia, and adenomatous lesions, as well as a marked reduction in gastric neoplasm miss rates from 27.3% to 6.1% during screening endoscopy[43,44]. In gastric motility assessment, an automated algorithm for artifact detection in body-surface gastric myoelectrical recordings achieved 95% accuracy, but evidence remains limited by the very small, all-female cohort[45].

Overall, the strongest evidence in gastric diseases is for AI-assisted detection of gastric neoplasia and precancerous lesions. Gastric cancer detection has been evaluated in large multicenter datasets with prospective and external validation, while AI-assisted prediction of submucosal invasion has also undergone prospective multicenter external validation[38,39]. Importantly, randomized clinical studies have shown reductions in gastric neoplasm miss rates and improved detection of intestinal metaplasia and adenomatous lesions, providing evidence of clinical workflow benefit beyond diagnostic accuracy[43,44]. However, most other applications, including H. pylori detection, gastric atrophy assessment, and gastric motility analysis, remain supported mainly by retrospective or small single-center studies. Heterogeneity in imaging platforms and endoscopist expertise limits direct comparison and generalizability of the study results. AI appears particularly beneficial for less-experienced endoscopists, although inappropriate reliance on poorly calibrated systems may also amplify diagnostic errors, underscoring the need for prospective validation across diverse settings and users. Studies evaluating AI applications in gastric diseases, including their methodologies, performance metrics, and key limitations, are summarized in Table 2. Figure 2 summarizes the key applications of AI in esophageal and gastric diseases.

Figure 2 Key studied and emerging applications of artificial intelligence in esophageal and gastric diseases: A conceptual summary. Bold and italic text indicates artificial intelligence applications supported by the strongest available clinical evidence. GERD: Gastroesophageal reflux disease; PSPW: Postreflux swallow-induced peristaltic wave; BE: Barrett’s esophagus; EC: Esophageal cancer; EHR: Electronic health record; POEM: Peroral endoscopic myotomy; FLIP: Functional luminal imaging probe; HREM: High-resolution esophageal manometry; H. pylori: Helicobacter pylori.

Table 2 Studies evaluating applications of artificial intelligence in gastric diseases.

Ref.	Clinical application	Data type	AI method	Sample size	Study design, center design, and validation design	Major performance metric	Principal limitation
Shichijo et al[32]	Detection of H. pylori infection	Endoscopic images (WLI)	22-layer deep CNN with GoogLeNet	32208 images (development), 11481 images from 397 patients (test)	Retrospective, single-center, independent test set comparison	Accuracy: 87.7%, 11481 images analyzed in approximately 3.2 minutes vs approximately 230 minutes by endoscopists	Single-center study
Liu et al[35]	Kyoto gastritis classification	Gastric endoscopy images (WLI and NBI)	Five GAM-EfficientNet models	29013 images from 2087 patients, test set of 528 images, compared with 10 endoscopists	Retrospective, single-center, internal training, validation, and testing sets	Average accuracy: 78.70%	Single-center study, imbalanced sample distribution, lack of paired histological validation
Hirasawa et al[36]	Gastric cancer detection	Endoscopic images (WLI, chromoendoscopy and NBI)	CNN with Single Shot MultiBox Detector	13584 training images, 2296 test images	Retrospective, multicenter training, independent test set of consecutive patients	Sensitivity: 92.2%, detected 100% of invasive cancers and 98.6% of lesions ≥ 6 mm, processed 2296 images in 47 seconds	Test set included only cancer cases with high disease prevalence, single endoscope type used
Ikenoyama et al[37]	Early gastric cancer detection	Endoscopic images (WLI)	CNN with Single Shot MultiBox Detector	13584 training images, 2940 test images	Retrospective multicenter training, single-center test set, independent comparison with 67 endoscopists	Sensitivity: 58.4% (95%CI: 51.4%-65.1%) in independent external test set, reading time 45.5 seconds vs 173 minutes for endoscopists	Test set limited to a single center and still images, low PPV of 26.0%
Luo et al[38]	Esophageal and gastric cancer detection	Endoscopy images (WLI)	Deep learning with GRAIDS	1036496 images from 84424 individuals	Multicenter case-control diagnostic study, internal, prospective, and external validation	Accuracy: 0.915-0.977 (external validation sets)	Only standard WLIs were used
Zhang et al[40]	Early gastric cancer detection	Gastroscopic images	Improved Mask R-CNN	1120 images of EGC (140 healthy, 980 EGC cases)	Retrospective, single-center, fivefold cross-validation and testing dataset	Accuracy: 93.9% (internal validation and testing dataset)	Small primary dataset from a single center, generalizability requires further verification
Bang et al[39]	Prediction of submucosal invasion in gastric neoplasms	WLI endoscopic images	Automated DL	2899 training images, 1597 prospective test images	Multicenter validation, training from one hospital and testing at another, prospective multicenter external test	Accuracy: 89.3% (95%CI: 85.1%-93.5%) (external validation), AI-assisted trainee accuracy 81.6% vs 68.4% without AI	Training images from a single institution
Gong et al[42]	Gastric neoplasm detection and classification	Real-time endoscopy	DL clinical decision support system	5017 training images, 2524 pilot procedures	Randomized pilot study and prospective multicenter external test	Classification accuracy: 81.5% (prospective multicenter external test)	Improvement in real-time lesion detection rate was not statistically significant
Wu et al[43]	Gastric neoplasm detection	Real-time white-light endoscopy	DL system	1812 patients	Single-center tandem RCT	Miss rate: 6.1% with AI vs 27.3% with routine examination	Single-center design, endoscopists were not blinded to randomization
Xu et al[44]	Detection of precancerous lesions and adenomas	Digestive endoscopy	ENDOANGEL	Prospective intervention phases	Prospective, randomized, single-center clinical trial	IM detection rate: 14.23% vs 9.15%	Single-center trial design
Calder et al[45]	Automated artifact detection and signal recovery in gastric body-surface mapping	High-resolution gastric myoelectrical activity data from BSGM	Automated artifact detection and rejection algorithm with signal recovery	10 patients, 176 data samples	Prospective validation study, single-center	Accuracy: 95%, sensitivity: 96%, specificity: 95% against expert markers	Small sample size and demographic bias, with all participants being female

H. pylori: Helicobacter pylori; WLI: White-light imaging; CNN: Convolutional neural network; NBI: Narrow-band imaging; GAM: Global attention mechanism; CI: Confidence interval; PPV: Positive predictive value; GRAIDS: Gastrointestinal artificial intelligence diagnostic system; EGC: Early gastric cancer; DL: Deep learning; AI: Artificial intelligence; RCT: Randomized controlled trial; IM: Intestinal metaplasia; BSGM: Body surface gastric mapping.

AI IN SMALL INTESTINAL DISEASES

AI has remarkably transformed small bowel diagnostics, particularly capsule endoscopy (CE), where it has significantly improved lesion detection, diagnostic accuracy, and workflow efficiency. DL models have demonstrated high accuracy for detecting diverse mucosal lesions, including tumors, ulcers, bleeding, and vascular lesions, with multicenter studies showing strong diagnostic performance[46-50]. However, most evidence remains retrospective and based on still images or selected frames, limiting validation in full-length CE videos. Notably, a multicenter study using images from seven CE platforms demonstrated robust cross-device performance, although external prospective validation remains needed[51].

In Crohn’s disease (CD), a prospective multicenter blinded study using full-length pan-enteric CE videos demonstrated 92%-96% sensitivity and 90%-93% specificity with prospective external validation, supporting a more clinically mature application of AI in CE[52]. A meta-analysis including 11 AI models has reported pooled sensitivity and specificity of 94% and 97%, respectively, with an area under the curve (AUC) of 0.99 for CD detection[53].

AI has also shown promising performance in celiac disease, including detection of duodenal villous atrophy on endoscopic and CE images and automated histopathological assessment, although most studies remain retrospective and subject to selection or small-sample limitations[54-57]. Beyond mucosal assessment, AI has been explored for intestinal motility and small-bowel obstruction. Preliminary studies demonstrated promising performance for identifying intestinal dysmotility and high-grade obstruction, but limited validation and retrospective designs constrain clinical generalizability[58,59].

Overall, the strongest evidence in small intestinal diseases is for AI-assisted CE, particularly automated detection of small-bowel mucosal lesions and CD. Large multicenter studies have demonstrated high diagnostic performance for automated lesion detection, while cross-device evaluation supports the potential for broader applicability[47,51]. AI-based detection of CD has also been validated prospectively in a multicenter blinded study using full-length pan-enteric capsule videos[52]. However, most other applications, including celiac disease assessment, intestinal dysmotility, and small-bowel obstruction, remain supported mainly by retrospective or small pilot studies. Reliance on expert consensus rather than independent clinical gold standards and the predominance of tertiary-center datasets further highlight the need for prospective, cross-device validation in diverse real-world populations. Studies evaluating AI applications in small intestinal diseases, including their methodologies, performance metrics, and key limitations, are summarized in Table 3.

Table 3 Studies evaluating applications of artificial intelligence in small intestinal diseases.

Ref.	Clinical application	Data type	AI method	Sample size	Study design, center design, and validation design	Major performance metrics	Principal limitation
Aoki et al[47]	Detection of various small intestinal mucosal lesions	CE images for training and videos for testing	Deep CNN	66028 images for training, 379 small-bowel videos for testing	Multicenter retrospective, independent test set of consecutive videos from 3 institutions, external validation	Per-patient detection rate: 99%, sensitivity for mucosal breaks 100% (95%CI: 88.4%-100%) and bleeding 100%, substantially reduced reading workload compared with conventional method	Per-patient assessment may not reflect detection of every individual small lesion
Ding et al[48]	Detection of various small intestinal mucosal lesions	SB-CE images and videos	ResNet-based CNN auxiliary model	Training: 158235 images/1970 patients, validation: 113426569 images/5000 patients	Multicenter retrospective, 77 medical centers, comparison of conventional and AI-assisted reading in 5000 patients, internal validation	Sensitivity: 99.88% (95%CI: 99.67%-99.96%) per patient and 99.90% (95%CI: 99.74%-99.97%) per lesion, reading time reduced from 96.6 ± 22.4 to 5.9 ± 2.1 minutes/patient, diagnostic yield increased from 74.57% to 99.88%	Retrospective design requires prospective evaluation
Liu et al[49]	Small-bowel tumor detection	Clinical patient video data	Multi-scale curvelet analysis and fractal technology with SVM	89 clinical patients	Retrospective, single-center context, feature extraction and classification testing on clinical video frames, internal validation	Sensitivity: 97.8%, specificity: 96.7%	Small dataset of only 89 patients
Leenhardt et al[50]	Detection of gastrointestinal angiectasia	SB-CE still frames	CNN with semantic segmentation	6360 frames from 4166 CE videos, 600-frame independent test set	Multicenter retrospective using a national database, two separate datasets for ML and algorithm testing, internal validation	Sensitivity: 100% (95%CI: 100%-100%), specificity: 96% (95%CI: 93.78%-98.22%)	Focused exclusively on typical GIA, performance on full-length videos not confirmed
Mascarenhas et al[51]	Detection of panendoscopic vascular lesions	Multi-brand and multi-device CE frames	RegNetY-based CNN	34655 frames from 1188 examinations	Multicenter retrospective cohort study, 5-fold cross-validation and independent exam-split test set	Sensitivity: 86.4%, specificity: 98.3%, accuracy: 95.0%, AUC 0.984 (95%CI: 0.977-0.991)	Retrospective selection bias, small number of frames in the testing dataset
Brodersen et al[52]	Detection of suspected CD	Pan-enteric CE videos	AXARO® DL platform	131 patients with suspected CD	Multicenter prospective, blinded trial, prospective external validation	Sensitivity: 92%-96%, specificity: 90%-93%, observer-specific AUC 0.91-0.94 (95%CIs: 0.86-0.98), 97.1% reduction in images requiring review, median reading time approximately 3.2 minutes/patient, 48.5% potentially ruled out without full video review	Relatively small patient cohort
Scheppach et al[54]	Detection of duodenal villous atrophy	Endoscopic still images	ResNet18 DL model	1704 images for internal validation, 349 images for external validation	Multicenter retrospective case-control, internal cross-validation and independent external test set	Sensitivity: 90%, specificity: 76%, accuracy: 84% (external validation)	High risk of selection bias due to retrospective case-control design
Stoleru et al[55]	Detection of celiac disease with villous atrophy	CE images	Linear SVM with image-processing kernels and filters	109 films, 100 frames per film, 51-film test set	Retrospective, single-center, comparison of Linear SVM, Fine KNN, and Weighted KNN, internal validation	Accuracy: 94.1%, sensitivity: 96%	Real-world clinical utility for full video streams was not the primary focus
Gruver et al[57]	Assessment of histologic features of celiac disease	Whole-slide images of IHC sections	DenseNet2 CNN classifiers	116 small-intestine biopsies, matched cohort of 28 patient pairs	Single-center retrospective pilot study, internal validation	More than 90% concordance with manual scores in matched biopsies	Small pilot study, manual selection of tissue by a pathologist remains necessary
Malagelada et al[58]	Identification of intestinal motor dysfunction	CE images	SVM	36 patients and 50 healthy subjects	Single-center prospective pilot study, leave-one-out cross-validation	100% identification of patients meeting manometric criteria for dysmotility	Complexity of parameters analyzed requires a much larger validation dataset
Cheng et al[59]	Detection of high-grade small-bowel obstruction	Abdominal radiographs	Inception-v3 CNN	7768 radiographs from 3212 patients, held-out test set of 2420 consecutive radiographs	Single-center retrospective study, two-stage training and evaluation on a held-out test set, internal validation	Sensitivity: 91.4%, specificity: 91.9%, ROC AUC: 0.971	Retrospective design, potential variation in ground truth as the second test set was assessed by a single radiologist

AI: Artificial intelligence; CE: Capsule endoscopy; CNN: Convolutional neural network; CI: Confidence interval; SB-CE: Small-bowel capsule endoscopy; SVM: Support vector machine; ML: Machine learning; GIA: Gastrointestinal angiectasia; AUC: Area under the curve; CD: Crohn’s disease; DL: Deep learning; KNN: K-nearest neighbors; IHC: Immunohistochemistry; ROC: Receiver operating characteristic; RegNetY: Regular network Y; ResNet: Residual network.

AI IN LARGE INTESTINAL DISEASES

Colorectal neoplasia

AI has transformed colorectal neoplasia management, particularly through computer-aided detection (CADe) systems that enable real-time polyp recognition during colonoscopy. Early studies established the feasibility of automated polyp detection[60]. Subsequent RCTs demonstrated improved adenoma detection rates with CADe, including multicenter evaluation of GI Genius, although benefits were smaller in some routine-practice settings[61-64]. A meta-analysis of 44 RCTs involving 36201 patients further confirmed increased adenoma detection rates and advanced neoplasia detection, with a modest increase in withdrawal time and non-neoplastic polyp resection[65].

Beyond lesion detection, AI-assisted computer-aided diagnosis (CADx) systems have shown high accuracy in real-time optical characterization of colorectal polyps, supporting “resect-and-discard” strategies and reducing the need for histopathological assessment for diminutive lesions[66]. However, a meta-analysis of 11 studies found no increase in the proportion of polyps avoiding pathology (90% vs 90%), despite a modest reduction in incorrect predictions (12% vs 13%; risk ratio = 0.88, 95% confidence interval: 0.79-0.98)[67].

AI has also been applied to digital pathology, with DL models accurately predicting microsatellite instability and mismatch repair deficiency from routine hematoxylin and eosin slides, supported by multicenter cohorts and external validation[68,69]. AI-based histopathological analysis has demonstrated prognostic value in colorectal cancer, improving risk stratification and survival prediction beyond conventional markers[70].

Inflammatory bowel disease

AI has played a substantial role in the diagnosis, disease activity assessment, prognostication, and personalized management of inflammatory bowel disease (IBD). An early ML study demonstrated improved differentiation of CD, ulcerative colitis (UC), and IBD-unclassified cases through the integration of endoscopic and histological features, although performance was lower for UC[71]. CNN- and endocytoscopy-based CAD systems have shown promising performance in assessing endoscopic and histologic activity in UC, with the latter achieving 91% accuracy for persistent histologic inflammation[72,73].

More recently, the Paddington International Virtual ChromoendoScopy ScOre (PICaSSO) showed strong correlation with histologic remission, and a PICaSSO of ≤ 3 was associated with lower rates of hospitalization, colectomy, and treatment escalation, outperforming conventional endoscopic scoring systems[74]. Building on this concept, the PICaSSO Histologic Remission Index and an accompanying CNN-based digital pathology platform were evaluated in a multicenter prospective study that automatically quantified histologic activity, differentiating active from quiescent disease with an accuracy of 86%[75].

Beyond disease activity assessment, ML models have shown potential for predicting IBD-related hospitalization and corticosteroid use, although generalizability was limited by the predominantly older, male population[76]. Using data from the UNITI and IM-UNITI trials, ML models predicted long-term remission in patients with CD receiving ustekinumab, achieving an AUC of 0.78 when early treatment-response data were incorporated[77].

Other diseases

AI has demonstrated promising applications in acute lower gastrointestinal bleeding (LGIB), diverticular disease, and irritable bowel syndrome. ANN and gradient-boosting models demonstrated good discrimination for predicting bleeding-related outcomes and treatment needs in LGIB, with external validation accuracies of 83%-97%, although model complexity limits clinical applicability[78,79]. Gradient boosting also predicted complicated acute diverticulitis with an external AUC of 0.85 and negative predictive value of 99%[80]. A random forest model developed in a prospective diagnostic study using gut microbiome and short-chain fatty acid data demonstrated excellent diagnostic performance for irritable bowel syndrome, with an AUC of 0.92[81].

Overall, the strongest clinical evidence in large intestinal diseases is for AI-assisted colorectal neoplasia detection, particularly real-time CADe during colonoscopy. Multiple randomized trials, including multicenter studies, and a meta-analysis of 44 RCTs have demonstrated improved adenoma detection, although the magnitude of benefit varies across clinical settings and may be accompanied by increased detection of non-neoplastic lesions and withdrawal time[61-65]. AI-assisted assessment of IBD is also progressing toward clinical application, with multicenter prospective evaluation of automated histologic activity assessment and evidence linking PICaSSO scores with clinically relevant outcomes such as hospitalization, colectomy, and treatment escalation[75]. Other applications, including CADx-based polyp characterization, digital pathology, IBD outcome prediction, and prediction of LGIB or diverticulitis, have shown promising diagnostic or prognostic performance but have less evidence of prospective clinical benefit. However, evidence for improved patient-centered outcomes and broader real-world implementation of most AI applications remains limited. Studies evaluating AI applications in large intestinal diseases, including their methodologies, performance metrics, and key limitations, are summarized in Table 4. Figure 3 summarizes the key applications of AI in small and large intestinal diseases.

Figure 3 Key studied and emerging applications of artificial intelligence in small and large intestinal diseases: A conceptual summary. Bold and italic text indicates artificial intelligence applications supported by the strongest available clinical evidence. CE: Capsule endoscopy; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CRC: Colorectal cancer; IBD: Inflammatory bowel disease; IBS: Irritable bowel syndrome; LGIB: Lower gastrointestinal bleeding.

Table 4 Studies evaluating applications of artificial intelligence in large intestinal diseases.

Ref.	Clinical application	Data type	AI method	Sample size	Study design, center design, and validation design	Major performance metrics	Principal limitation
Misawa et al[60]	Colorectal polyp detection	Colonoscopy videos	Convolutional 3D neural network	73 patients, learning set: 411 short videos/59 patients, independent test set: 135 short videos/14 patients	Single-center retrospective, independent test set, internal validation	Sensitivity 90.0%, specificity 63.3%	High false-positive rate (60%) in polyp-based analysis
Wang et al[63]	Polyp and adenoma detection	Real-time colonoscopy video stream	DL with SegNet-based DCNN	1058 patients, 522 CADe vs 536 control	Single-center RCT	ADR 29.1% vs 20.3% in control, OR 1.61 (95%CI: 1.213-2.135)	Single-center study, inability to blind endoscopists
Gong et al[64]	Adenoma detection and technical quality	Real-time colonoscopy videos	ENDOANGEL with DCNNs and perceptual hash algorithms	704 patients (355 ENDOANGEL vs 349 control)	Single-center RCT	ADR 16% vs 8% in control, OR 2.30 (95%CI: 1.40-3.77)	Single-center study, artificially defined speed categories
Repici et al[61]	Colorectal neoplasia detection	Real-time colonoscopy images	DL with GI-Genius	685 subjects (341 CADe vs 344 control), AI training set: 2684 polyps from 840 patients	Multicenter RCT	ADR 54.8% vs 40.4% in control, RR 1.30 (95%CI: 1.14-1.45)	Expert setting with experienced operators, potential psychological bias
Karsenti et al[62]	Colorectal adenoma detection	Routine real-time colonoscopy	Real-time CADe with GI Genius 2.0.2	2015 participants (1003 CADe vs 1012 control)	Single-center RCT	ADR 37.5% vs 33.7% in control, absolute difference 4.1 percentage points (95%CI: 0.0-8.1, P = 0.051)	Primary ADR outcome reached only borderline statistical significance
Byrne et al[66]	Polyp differentiation, adenoma vs hyperplastic polyp	Unaltered NBI colonoscopy video frames	DCNN with Inception	125 diminutive polyps	Retrospective single-center expert-operator videos, independent test set, internal validation	Accuracy 94% (95%CI: 86-97), sensitivity 98% (95%CI: 92-100) for adenomas	Videos collected by a single expert operator, not tested on live patients
Kather et al[68]	Microsatellite instability prediction	H&E-stained histology slides	Deep residual learning with ResNet18	Various cohorts (discovery: 1380, external: 378)	Multicenter discovery and validation, external validation	Patient-level AUC 0.84 (95%CI: 0.72-0.92) for colorectal cancer	Performance may not generalize beyond cancer types or ethnicities represented in training
Echle et al[69]	Microsatellite instability prediction	Routine H&E histology slides	Modified ShuffleNet DL	8836 colorectal tumors, 6406 development, 771 external resection validation	International multicenter retrospective, external validation	AUROC 0.96 for dMMR	Performance significantly lower in small biopsy samples than resections
Skrede et al[70]	Colorectal cancer prognosis	Scanned conventional H&E sections	Integration of 10 CNNs	2473 training and tuning, 2042 validation data	Retrospective discovery and validation, independent external validation	Hazard ratio 3.84 (95%CI: 2.72-5.43; P < 0.0001) for poor vs good prognosis	Retrospective cohorts, clinical utility requires prospective evaluation
Mossotto et al[71]	Pediatric IBD classification, CD, UC and IBDU	Endoscopic and histological findings from 10 GI locations	Linear SVM with RFE-CV	287 children, 239-patient model development/testing cohort and independent validation of 48 patients	Retrospective single-center, internal and independent validation	Accuracy 82.7% combined model, 83.3% independent validation, IBDU cases were reclassified as CD or UC with > 80% posterior probability	Less accurate for UC (65%) than CD (94%)
Ozawa et al[72]	UC endoscopic activity assessment	Colonoscopy images	CNN with GoogLeNet	26304 images from 841 patients for training, 3981 images from 114 patients for testing	Retrospective image analysis, internal, independent image test set validation	AUROC 0.86 for Mayo 0 and 0.98 for Mayo 0-1	Performance varied by location, lower accuracy in right and left colon than rectum
Maeda et al[73]	UC histologic activity assessment	Endocytoscopy images at 520 × ultra-magnification	SVM using 312 extracted texture features	22835 images from 187 patients, training 12900 images/87 patients, validation 9935 images from 100 patients	Retrospective validation, single-center, temporal training and validation split, internal validation	Accuracy 91% (95%CI: 88-93), sensitivity 74% (95%CI: 65-81), specificity 97% (95%CI: 95-99)	Images collected by experts only, no comparison with long-term clinical prognosis
Gui et al[75]	UC histologic activity assessment	Histological whole-slide images	CNN	138 biopsies from 307 patients	Prospective multicenter, 11 centers, 70% training and 30% testing, internal validation	Accuracy 86%, sensitivity 78%, specificity 91.7%	Lost predictive power in patients who had achieved endoscopic remission
Waljee et al[76]	IBD outcome prediction, UC, CD and indeterminate colitis	Longitudinal laboratory data and demographics from EMR	Random forest	20368 patients	Retrospective cohort, National VHA database, temporal 70% training and 30% testing split, internal validation	AUROC 0.85 (95%CI: 0.84-0.85) to 0.87 (95%CI: 0.87-0.88) depending on model	Older predominantly male population may limit generalizability
Waljee et al[77]	CD outcome prediction	Demographic and laboratory data from clinical trials	Random forest	401 participants, repeated 70/30 training-testing splits	Cohort study of three phase 3 RCTs, repeated 70% and 30% training and testing splits, internal validation	AUROC 0.78 (95%CI: 0.69-0.87) for week-8 model	CRP is an imperfect proxy for remission
Das et al[78]	Acute lower gastrointestinal bleeding risk prediction	Non-endoscopic clinical variables	ANN	332 patients, 120 training, 70 internal validation, 142 external validation	Prospectively collected validation study, multicenter, internal and independent external validation	Accuracy 87%-96% (internal validation), 93%-97% (external validation)	Low PPV for mortality (31%-63%), 26 variables are more cumbersome
Ayaru et al[79]	Acute lower gastrointestinal bleeding risk prediction	Non-endoscopic variables	Gradient boosting	300 patients, 170 for training and internal validation, 130 for external validation	Retrospective database study, multicenter, internal and independent external validation	Accuracy 78%-88% (internal validation), 83%-91% (external validation)	Model requires 39 input variables, mortality was not examined because of its rarity
Klang et al[80]	Acute diverticulitis prediction	EMR data	Gradient boosting	4997 visits, 3154 for internal validation, 1343 for external validation	Multicenter retrospective cohort, internal and held-out external validation	AUC 0.85 (95%CI: 0.78-0.91), sensitivity 88%, NPV 99% (external validation)	Retrospective design and missing data, including CRP missing in 95% of patients
Fukui et al[81]	Irritable bowel syndrome diagnosis	Fecal gut microbiome profiles using 16S rRNA and short-chain fatty acids	Random forest using LASSO feature extraction	111, 85 IBS patients and 26 healthy controls	Prospective diagnostic study, multicenter secondary and tertiary care clinics, internal repeated 10-fold cross-validation	AUC 0.911-0.920, sensitivity > 80%, specificity > 90%	Small control group, lack of dietary information and functional meta-omics investigations

AI: Artificial intelligence; DL: Deep learning; DCNN: Deep convolutional neural network; CADe: Computer-aided detection; RCT: Randomized controlled trial; ADR: Adenoma detection rate; OR: Odds ratio; GI: Gastrointestinal; RR: Relative risk; NBI: Narrow-band imaging; CI: Confidence interval; H&E: Hematoxylin and eosin; AUC: Area under the curve; AUROC: Area under the receiver operating characteristic curve; dMMR: Deficient mismatch repair; CNN: Convolutional neural network; IBD: Inflammatory bowel disease; CD: Crohn’s disease; UC: Ulcerative colitis; IBDU: Inflammatory bowel disease unclassified; SVM: Support vector machine; RFE-CV: Recursive feature elimination with cross-validation; EMR: Electronic medical record; VHA: Veterans Health Administration; CRP: C-reactive protein; ANN: Artificial neural network; PPV: Positive predictive value; NPV: Negative predictive value; IBS: Irritable bowel syndrome; LASSO: Least absolute shrinkage and selection operator; ResNet: Residual network; rRNA: Ribosomal ribonucleic acid; SegNet: Semantic segmentation network.

AI IN PANCREATIC DISEASES

AI has demonstrated considerable potential across the spectrum of pancreatic diseases, including pancreatic cancer, cystic neoplasms, and acute pancreatitis. A radiomics-based ML model detected pancreatic ductal adenocarcinoma (PDAC) on prediagnostic computed tomography (CT) scans with 92.2% accuracy and 95.5% sensitivity[82]. The Radiomics-based Early Detection model subsequently demonstrated detection of visually occult PDAC on routine CT imaging in a multicenter development and validation study, while the PANORAMA study further validated AI-assisted PDAC detection across multinational cohorts[83,84]. Apart from imaging, CancerRiskNet, a DL model using longitudinal EHRs, accurately predicted pancreatic cancer risk up to three years before diagnosis, although performance declined across different healthcare systems[85]. AI-assisted endoscopic ultrasound (EUS) has also shown high diagnostic performance for pancreatic malignancy, but evidence remains limited by small study numbers and predominantly offline analysis[86,87].

AI has also shown substantial utility in pancreatic cystic neoplasms. A DL model based on EUS images of intraductal papillary mucinous neoplasms achieved 94.0% accuracy for predicting malignancy, although the small, single-center cohort comprising only surgical cases limits generalizability[88]. Similarly, a DL model integrating cyst fluid biomarkers, clinical variables, and imaging characteristics achieved excellent performance for differentiating malignant from benign cystic lesions, surpassing conventional biomarkers and cytology[89]. DL models based on CT imaging and needle-based confocal laser endomicroscopy have further improved the classification and risk stratification of pancreatic cystic lesions, simultaneously enhancing radiologist performance and reducing interpretation time[90-92]. AI has also been explored for predicting acute pancreatitis severity and complications, but most evidence remains retrospective and lacks external validation[93].

Overall, the strongest evidence in pancreatic diseases is emerging for AI-assisted detection of pancreatic cancer and risk stratification of pancreatic cystic neoplasms. PDAC detection has progressed from retrospective models to international multicenter external validation, while AI-based cystic lesion assessment has also undergone multicenter external and prospective validation[83,84,91]. These applications show greater clinical maturity than EUS-based pancreatic cancer diagnosis, intraductal papillary mucinous neoplasm classification, or acute pancreatitis prediction, which remain largely limited by small cohorts, retrospective designs, or limited external validation. Although some models have demonstrated potential to improve diagnostic performance and reduce interpretation time, evidence of improved clinical decision-making or patient outcomes remains limited. Studies evaluating AI applications in pancreatic diseases, including their methodologies, performance metrics, and key limitations, are summarized in Table 5.

Table 5 Studies evaluating applications of artificial intelligence in pancreatic diseases.

Ref.	Clinical application	Data type	AI method	Sample size	Study design, center design, and validation design	Major performance metric	Principal limitation
Mukherjee et al[82]	PDAC prediction	Prediagnostic CT scans	SVM	420 CT scans, 155 prediagnostic, 265 controls	Retrospective case-control, internal test set, independent internal and external NIH sets, internal and external validation	AUC 0.98 (95%CI: 0.94-0.98), accuracy 92.2% (internal validation), specificity 96.2% (external validation), outperformed radiologists (mean AUC 0.66)	Retrospective design, small external validation cohorts
Mukherjee et al[83]	PDAC prediction	Standard-of-care CT scans	Ensemble with logistic regression, random forest, and XGBoost	1462 scans, 969 training and 493 test	Retrospective, multicenter, independent test set and external specificity validation cohorts	AUC 0.82 (95%CI: 0.81-0.83), sensitivity 73.0%, specificity 81.1% (internal validation), 81.3% specificity (external validation)	Retrospective design, prospective validation in high-risk cohorts is required
Alves et al[84]	PDAC detection	Routine CT scans	DL	3440 patients, 2224 for training, 86 for tuning, 1130 for sequestered test, multi-reader substudy of 391 cases assessed by 68 radiologists	International multicenter observational study, sequestered testing cohort and radiologist subset	Mean AUROC 0.92 (95%CI: 0.90-0.93) (external validation), AI performance was statistically superior to the 68-radiologist comparison (P < 0.001)	Observational design, results represent average radiologist performance
Placido et al[85]	Pancreatic cancer prediction	Health records with ICD codes	Transformer and GRU	9.2 million patients from DNPR and US-VA	Retrospective cohort, random 10% test split and cross-application between health systems, internal and cross-validation	AUROC 0.88 in Denmark and 0.78 in US-VA (internal validation), Danish model on US-VA dataset yielded AUROC 0.71	Performance declined when applied across different healthcare systems
Kuwahara et al[88]	IPMN classification	EUS digital still images	DL with ResNet50	50 patients, 3970 images	Retrospective single-center, 10-fold cross-validation	Accuracy 94.0%, AUC 0.98, diagnostic accuracy 94.0% vs 56.0% for human preoperative diagnosis	Small sample size, retrospective design, only surgical cases included
Kurita et al[89]	Pancreatic cystic lesion classification	Cyst fluid markers and clinical data	DL with TensorFlow	85 patients	Retrospective, single-center, observational study, 5-fold cross-validation	AUROC 0.966, accuracy 92.9%, AI outperformed cyst-fluid carcinoembryonic antigen (AUROC 0.719) and cytology (AUROC 0.739)	Small sample size, single-center study, subjective input factors
Machicado et al[90]	IPMN classification	nCLE video frames	CNN	35 patients, 15027 frames	Post hoc analysis of a prospective multicenter cohort, performance on post hoc video data, internal validation	Accuracy 82.9%-85.7%, sensitivity 83.3%	Small patient cohort, limited to pathologically confirmed cases
Yuan et al[91]	Pancreatic cystic neoplasm malignancy prediction	Contrast-enhanced CT scans	Mamba U-Net and XGBoost	1835 patients, 1249 primary cohort (999 training, 250 validation), external test sets of 173, 289 and 124 patients	Multicenter retrospective and prospective study, external test sets and prospective real-world cohort, internal and external validation	AUC 0.845-0.988 (external validation), malignancy detection 81.8% (prospective validation), interpretation time reduced 23.7%, and workload reduced 39.3%	Focused on intralesional features, did not incorporate peripancreatic findings
Duh et al[92]	Pancreatic cystic lesion classification	Abdominal CT scans	Attention Gate U-Net	335 abdominal CT scans, 297 pancreatic cystic lesions and 38 controls, training 241 scans/93 patients, independent test 94 scans/43 patients	Retrospective single-center, independent testing set of 94 scans, internal validation	Sensitivity 93.1%, specificity 81.8%	Single-center design, small control cohort, lacks external validation

AI: Artificial intelligence; PDAC: Pancreatic ductal adenocarcinoma; CT: Computed tomography; SVM: Support vector machine; NIH: National Institutes of Health; AUC: Area under the curve; CI: Confidence interval; XGBoost: Extreme gradient boosting; DL: Deep learning; AUROC: Area under the receiver operating characteristic curve; ICD: International Classification of Diseases; GRU: Gated recurrent unit; DNPR: Danish National Patient Registry; US-VA: United States Veterans Affairs; IPMN: Intraductal papillary mucinous neoplasm; EUS: Endoscopic ultrasonography; nCLE: Needle-based confocal laser endomicroscopy; CNN: Convolutional neural network; ResNet: Residual network.

AI IN BILIARY DISEASES

AI has emerged as a valuable tool in the diagnosis, risk stratification, and endoscopic management of biliary disorders. The ModelArts prediction system significantly outperformed conventional risk stratification criteria for choledocholithiasis, achieving an accuracy of 97% in a retrospective study, while a meta-analysis confirmed high diagnostic performance of AI-assisted common bile duct stone prediction[94,95].

ML models using routine clinical biomarkers have accurately differentiated benign from malignant distal bile duct obstruction in a single-center retrospective study[96]. CNN-based analysis of digital cholangioscopy images also demonstrated high performance for differentiating benign from malignant biliary strictures, with subsequent multicenter validation showing promising performance, although evaluation remained retrospective and platform-specific[97-100].

AI applications in therapeutic endoscopy have mainly focused on improving endoscopic retrograde cholangiopancreatography (ERCP) outcomes. In a prospective pilot study, an ML model was developed using pooled data from 12 RCTs that accurately predicted post-ERCP pancreatitis (PEP), achieving a negative predictive value of 95% and supporting individualized risk assessment and prophylaxis selection[101]. However, the positive predictive value (PPV) was low, and the model was built on trial data, which may differ from real-world populations. Several ML and radiomics-based models have shown improved prediction of PEP and post-ERCP cholangitis, outperformed conventional statistical approaches, and enabled more precise risk stratification[102-104]. AI-based ampulla detection and cannulation difficulty prediction is another emerging application, although recall for difficult cases remains low[105].

The strongest evidence in biliary diseases is for AI-assisted prediction of choledocholithiasis and characterization of biliary strictures, although most studies remain retrospective and require broader external validation[94-100]. AI models for PEP have progressed to prospective evaluation, but their predictive performance remains modest and prospective validation has been limited[101,102,104]. Further, low complication rates in PEP and cholangitis prediction models can inflate negative predictive values while limiting positive predictive performance. AI-assisted ampulla detection and prediction of difficult cannulation are still early-stage applications[105]. Overall, AI in biliary disease remains largely at the diagnostic and risk-prediction stage, with limited evidence of improved clinical workflow or patient outcomes and little evidence of real-world implementation. Studies evaluating AI applications in biliary diseases, including their methodologies, performance metrics, and key limitations, are summarized in Table 6.

Table 6 Studies evaluating applications of artificial intelligence in biliary diseases.

Ref.	Clinical application	Data type	AI method	Sample size	Study design, center design, and validation design	Major performance metric	Principal limitation
Zhang et al[94]	Choledocholithiasis prediction	Preoperative clinical predictors	ModelArts	1199 patients	Retrospective, single-center, internal validation	Accuracy 0.97, recall 0.97, precision 0.971	Retrospective design may limit generalizability to real-world settings
Sugimoto et al[96]	Distal bile duct obstruction, benign vs malignant	Clinical laboratory biomarkers	LightGBM	206 patients, 64 malignant, 142 benign, 16 excluded for missing values	Retrospective, single-center, internal validation with 10-fold cross-validation	AUROC 0.908, sensitivity 83.1%, specificity 87.2%, AUROC significantly higher than individual biomarkers	Small sample size, exclusion of markers such as CA19-9 because of missing data
Saraiva et al[99]	Biliary stricture, benign vs malignant	Digital single-operator cholangioscopy images	CNN	85 patients, 11855 images	Retrospective pilot study, single-center, internal validation with 5-fold cross-validation	Accuracy 94.9%, sensitivity 94.7%, specificity 92.1%, AUC 0.988	Small patient cohort, static images rather than full videos
Saraiva et al[100]	Malignant biliary stricture and morphologic characterization	Digital single-operator cholangioscopy images	ResNet-based CNN	129 DSOC examinations, 84994 images	Retrospective, multicenter, internal validation	Overall accuracy 82.9%, AUROC 0.92, tumor vessel detection AUC 0.98, papillary projection detection AUC 0.96	Tested on a single DSOC platform, technological maturity remains preclinical
Brenner et al[101]	Post-ERCP pancreatitis prediction	Patient-level data from 12 RCTs	Gradient-boosted machine	7389 patients for development, 135 patients for pilot	Retrospective development and prospective pilot, multicenter development and single-center pilot, internal and prospective validation	AUC 0.70 (internal cross-validation), AUC 0.74 (prospective pilot)	Low PPV of 15%, trial data may not reflect real-world populations
Archibugi et al[102]	Post-ERCP pancreatitis prediction	Clinical preprocedural and procedural factors	Gradient boosting	1150 patients	Prospective multicenter international cohort, proof-of-concept, internal validation	AUC 0.70 (cross-validation), AUC 0.671 (internal test set)	Moderate predictive accuracy in the final test set
Jin et al[103]	Post-ERCP cholangitis prediction in malignant biliary obstruction	Clinical laboratory and procedural data	XGBoost	1421 patients, derivation cohort 1026 (718 training, 308 testing), external validation 395	Retrospective, multicenter, internal and external validation	AUC 0.7670 (internal test), AUC 0.7270 (external validation cohort)	Retrospective design, external validation cohort restricted to the same geographic region
Chen et al[104]	Post-ERCP pancreatitis prediction	Papilla radiomics from endoscopic images and clinical data	DL and ML algorithms	2372 patients	Retrospective multicohort, multicenter, internal validation	AUC 0.825-0.857 (internal validation and testing cohorts)	Retrospective enrollment, focus on surface morphology rather than internal structural factors
Kim et al[105]	Ampulla detection and cannulation difficulty prediction	Endoscopic images	CNN with U-Net and ResNet	531 patients for difficulty, 451 for detection	Retrospective, multicenter, internal validation with 5-fold cross-validation	mIoU 64.1% for detection, recall 61.1% for difficult cases	Low recall for difficult cases, lack of radiologic data to capture internal ampullary structure

AI: Artificial intelligence; AUROC: Area under the receiver operating characteristic curve; CA19-9: Carbohydrate antigen 19-9; CNN: Convolutional neural network; AUC: Area under the curve; DSOC: Digital single-operator cholangioscopy; ERCP: Endoscopic retrograde cholangiopancreatography; XGBoost: Extreme gradient boosting; DL: Deep learning; ML: Machine learning; mIoU: Mean intersection over union; LightGBM: Light Gradient Boosting Machine; PPV: Positive predictive value; RCT: Randomized controlled trial; ResNet: Residual neural network; U-Net: U-shaped convolutional neural network.

AI IN LIVER DISEASES

AI has emerged as a valuable tool across the spectrum of liver diseases, assisting disease detection, fibrosis staging, and differential diagnosis. In metabolic dysfunction-associated steatotic liver disease (MASLD), ML models developed using 17 demographic and clinical variables outperformed conventional non-invasive tests, such as FibroScan and FIB-4, for identifying clinically significant fibrosis and cirrhosis[106]. SMART AI-PATHO is a fully automated system for assessing steatosis and fibrosis in MASLD liver biopsy specimens, demonstrating 68.4% overall agreement with pathologist-assigned Metavir scores, increasing to 80.8% for significant fibrosis (F2-F4) and 89.7% for advanced fibrosis (F3-F4)[107].

A DL-based radiologist complementarity decision system using portal venous CT achieved AUCs of 0.89-0.92, an Obuchowski index of 0.92 internally, and 0.81-0.85 externally[108]. Radiomics and DL models using magnetic resonance imaging, ultrasound elastography, and automated image analysis also demonstrated high performance for fibrosis staging, in some cases outperforming conventional imaging and serum biomarkers[109-112].

AI has also shown high performance in differentiating benign from malignant focal liver lesions in a large multicenter cohort with external validation[113]. An explainable ML model based on clinical and laboratory data has shown strong performance in identifying viral hepatitis, although lacking validation on real-world clinical datasets[114]. AI models have successfully distinguished clinically challenging conditions with overlapping features such as drug-induced liver injury from autoimmune hepatitis (AIH), AIH from primary biliary cholangitis, and alcohol-associated liver disease from MASLD[115-118].

DL-based pathological phenotyping has enabled differentiation of hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma, and combined hepatocellular-cholangiocarcinoma, although the absence of an absolute ground truth for biphenotypic reclassification remains a limitation[119].

An ML analysis of clinical and metabolomic data derived a metabolome-based score that accurately identified metabolic dysfunction-associated steatohepatitis and strongly predicted liver-related mortality, with validation across independent cohorts[120]. Likewise, metabolomics and ML-based models have shown robust performance in predicting cirrhosis-related complications and identifying individuals at high risk for progressive liver disease, although some rely on non-routinely available biomarkers[121,122].

Imaging and histopathology-based models have demonstrated the ability to predict hepatic decompensation, HCC, and mortality more accurately than conventional risk scores, with validation across multicenter cohorts[123]. A DL model using digitized liver biopsies predicted future HCC development and identified high-risk patients, including those with mild fibrosis[124]. Complementing these approaches, a transcriptomic ML model has shown promise for the early detection of HCC across independent patient cohorts[125].

AI applications have extended to liver transplantation. GraftIQ, a hybrid neural network integrating clinicopathological data and expert clinician knowledge, non-invasively classifies post-transplant graft pathology. Validated across international centers, it outperformed both standard ML models and experienced hepatologists in identifying common causes of graft dysfunction[126].

The evidence in liver diseases is strongest for AI-assisted fibrosis staging, focal liver lesion characterization, and prediction of clinically relevant liver outcomes. Fibrosis assessment has progressed to multicenter studies with external or hospital-cross validation, including prospective evaluation in hepatitis B virus-related fibrosis, while focal liver lesion characterization has demonstrated strong performance in a large multicenter cohort with external-center validation[108,111-113]. AI-based prediction of drug-induced liver injury vs AIH and pathological classification of primary liver malignancies have also undergone external and, in some cases, prospective validation[118,119]. Prognostic applications, including prediction of metabolic dysfunction-associated steatohepatitis, cirrhosis-related complications, hepatic decompensation, and HCC development, show promising performance across independent cohorts, although several models depend on specialized biomarkers, imaging modalities, or retrospective datasets[120-125]. AI applications in liver transplantation have also reached multicenter external validation, as demonstrated by GraftIQ for post-transplant graft pathology[126]. However, cross-study performance is difficult to compare because reference standards, disease prevalence, patient selection, and imaging or pathology platforms vary, limiting model generalizability across populations and healthcare settings. Models requiring advanced imaging, metabolomics, or digital pathology may also face greater barriers to deployment in resource-limited settings. Overall, evidence remains dominated by diagnostic and prognostic performance, with limited demonstration of improved clinical workflow, patient-centered outcomes, or prospective real-world implementation.

Studies evaluating AI applications in liver diseases, including their methodologies, performance metrics, and key limitations, are summarized in Table 7. Figure 4 summarizes the key applications of AI in pancreatic, biliary, and liver diseases. Table 8 summarizes the major AI applications in gastroenterology according to clinical setting, evidence level, external validation, regulatory status, and key unresolved limitations.

Figure 4 Key studied and emerging applications of artificial intelligence in pancreatic, biliary, and liver diseases: A conceptual summary. Bold and italic text indicates artificial intelligence applications supported by the strongest available clinical evidence. MASLD: Metabolic dysfunction-associated steatotic liver disease; MASH: Metabolic dysfunction-associated steatohepatitis; HCC: Hepatocellular carcinoma; ERCP: Endoscopic retrograde cholangiopancreatography; EHR: Electronic health record; PDAC: Pancreatic ductal adenocarcinoma.

Table 7 Studies evaluating applications of artificial intelligence in liver diseases.

Ref.	Clinical application	Data type	AI method	Sample size	Study design, center design, and validation design	Major performance metric	Principal limitation
Chang et al[106]	NAFLD fibrosis assessment	Demographic and clinical features	Random forest, logistic regression, and ANN	1370 patients, training 1096 and testing 274	Multicenter cohort, internal validation	AUC 0.89 for cirrhosis	Requires further clinical validation for widespread use
Preechathammawong et al[107]	NAFLD fibrosis and steatosis assessment	Digital histopathological images, H&E biopsy	SMART AI-PATHO, DL algorithms	146 hepatic tissue specimens	Cross-sectional analytic, multicenter observational, comparison with pathologist interpretations	Agreement 68.4%, kappa 0.59 for all staging, 89.7% agreement for advanced fibrosis	Cannot assess lobular inflammation, ballooning, or variations in staining techniques
Zheng et al[108]	Liver fibrosis staging	Portal venous CT imaging	DRCDS, DL segmentation and classification with decision model	819 patients in the development set and 175 patients in external test sets	Retrospective multicenter study, internal test set and 3 external test sets	AUC 0.89-0.92 (internal test), Obuchowski index 0.81-0.85 (external validation), reduced radiologist evaluation requirement by 73.7%-92.0% while maintaining/improving sensitivity and specificity	Retrospective design, potential sampling errors in the biopsy reference standard
Xue et al[109]	Liver fibrosis assessment	Grayscale and elastogram ultrasound images	Transfer learning with Inception-V3	466 patients undergoing partial hepatectomy, training 364 and test 102	Retrospective diagnostic study, single-center, random-split internal test set	AUC 0.950 (95%CI: 0.917-0.972) for Scheuer stage S4	Uneven distribution of fibrosis stages
Ai et al[110]	Liver fibrosis assessment	Ultrasound radiofrequency signals and B-mode images	2D CNN for ROI analysis and 1D CNN for classification	613 participants for ROI, 237 participants for classification	Feasibility study, single-center, internal test sets	AUC 0.957 for ≥ METAVIR F1 (phase-spectrum test set)	Limited sample size and single-scanner diversity
Zha et al[111]	Liver fibrosis assessment	MRI and clinical variables	ResUNet for segmentation and CoRC logistic regression model	595 patients with biopsy-proven liver fibrosis	Retrospective multicenter cohort, internal and external validation	AUC 0.79-0.82 (internal and external test sets)	Retrospective study design
Wang et al[112]	Liver fibrosis in HBV	2D shear-wave elastography	DL radiomics with CNN	398 patients, training 266 and validation 132	Prospective multicenter study, internal and hospital-cross validation	AUC 0.97 (95%CI: 0.94-0.99) for METAVIR F4 (validation set)	Accuracy sensitive to the number of acquisitions
Wei et al[113]	Focal liver lesion characterization	Multiphase enhanced CT imaging	Liver Lesion Network	4039 patients	Multicenter development study, external-center validation, and clinical verification	AUC 97.2% for benign vs malignant lesions (external centers)	Primarily intended to aid areas with limited radiologist availability
Sharma et al[114]	Hepatitis C detection	Standard biochemical blood parameters	Stacked meta-model with random forest and XGBoost, with pre-clustering	869 patients	Hybrid dataset from NHANES and UCI, multi-source, holdout validation	Accuracy 94.82% (holdout set)	Lacks validation on real-world datasets
Ercan et al[115]	Autoimmune hepatitis assessment	H&E and Sirius red biopsy slides	CNN pipeline for segmentation and detection	123 pairs of biopsy slides (H&E and Sirius red) from 116 patients	Diagnostic tool development, single-center, independent 20% test set	Accuracy 99.4% for tissue detection (test set)	Relies solely on images without clinical data
Gerussi et al[116]	Autoimmune hepatitis vs primary biliary cholangitis	H&E digitized biopsy slides	Transformer-based attention-based multiple instance learning with ALNE	446 patients providing a dataset of 561 digitized whole-slide images	Multicenter diagnostic study, internal cross-validation and external validation	AUC 0.81 (external validation)	Limited sample size and focus on Western Europe
Park et al[117]	Alcohol-associated liver disease vs metabolic dysfunction-associated steatotic liver disease	Fecal 16S rRNA sequencing data	CNN with principal component analysis	674 subjects (464 for internal validation and 210 for external validation)	Multicenter screening signature development, internal cross-validation and external validation	AUC > 0.90 for alcohol-associated liver disease (external validation)	Metagenomic signatures are not universally established
Wang et al[118]	Drug-induced liver injury vs autoimmune hepatitis	24 routine laboratory parameters	Gradient boosting decision tree	2554 patients	Multicenter cohort, external and prospective validation	AUROC 0.91 (95%CI: 0.900-0.928) (external validation)	Web-based tool requires clinical implementation
Calderaro et al[119]	HCC vs intrahepatic cholangiocarcinoma and combined hepatocellular-cholangiocarcinoma	H&E whole-slide images	Self-supervised ResNet50 with attention-based multiple instance learning	1356 cases	Multicenter phenotyping and reclassification study, internal cross-validation and external TCGA validation	AUROC 0.94 (external TCGA validation)	No ground truth for biphenotypic reclassification
Huang et al[120]	Metabolic dysfunction-associated steatohepatitis	44 clinical variables and 250 serum metabolites	Multi-step ML	616 patients (311 in the derivation cohort and 305 in the validation cohort)	Multicenter development study across different countries, external validation in Finnish cohort	AUROC 0.87 (derivation cohort)	Requires serum metabolite measurements by NMR
Guo et al[121]	Prediction of cirrhosis complications	168 circulating metabolites and clinical variables	Integrated nomogram	64005 participants from the UK Biobank, training 43734, validation 20271	Multicenter UK Biobank tool-development study, internal and prospective evaluation	AUC 0.930 for 3-year risk (training and validation cohort)	Relies on metabolites not routinely measured
Njei et al[122]	High-risk metabolic dysfunction-associated steatotic liver disease	Demographic, clinical and laboratory data	XGBoost with SHAP	5281 subjects from the NHANES database	Multicenter observational national survey, internal holdout validation	AUC 0.95 (95%CI: 0.91-0.97) (holdout test set)	Lacked liver biopsy data for comparison
Kim et al[123]	Prediction of hepatic decompensation	MRE stiffness and clinical variables	Multivariable Cox regression	1254 participants with NAFLD from an international cohort	Multicenter cohort, internal random-split validation	C-statistic 0.876 for 5-year risk (validation cohort)	Requires MRE and has retrospective design
Nakatsuka et al[124]	HCC prediction in steatotic liver disease	H&E whole-slide images	CNN with EfficientNetB0	685 patients	Retrospective multicenter prediction study, external validation using nonpaired cases	AUC 0.84 (95%CI: 0.72-0.96) (external validation cases)	Retrospective design and focus on a Japanese population
Zhang et al[125]	Early HCC prediction	Gene expression profiles using secreted genes and relative expression ordering	Hybrid model with relative expression ordering, feature selection, and SVM	5586 tissue samples	ML signature development using public databases, independent validation sets	Accuracy 98.34% (independent validation)	Primarily based on public databases
Sharma et al[126]	Graft injury after liver transplantation	Demographic, clinical and laboratory data	GraftIQ with multiclass neural network and data fusion	5759 biopsies	Multicenter diagnostic study, internal test set and external validation	AUC 0.934 (95%CI: 0.909-0.959) (external UHN validation set)	Exclusion of biopsies with dual diagnoses

AI: Artificial intelligence; NAFLD: Nonalcoholic fatty liver disease; ANN: Artificial neural network; AUC: Area under the curve; H&E: Hematoxylin and eosin; DL: Deep learning; CT: Computed tomography; DRCDS: Deep learning-radiologist complementarity decision system; CI: Confidence interval; CNN: Convolutional neural network; ROI: Region of interest; MRI: Magnetic resonance imaging; ResUNet: Residual U-Net; CoRC: Concordance receiver operating characteristic; HBV: Hepatitis B virus; XGBoost: Extreme gradient boosting; NHANES: National Health and Nutrition Examination Survey; UCI: University of California, Irvine; ALNE: Autoimmune liver neural estimator; rRNA: Ribosomal ribonucleic acid; AUROC: Area under the receiver operating characteristic curve; HCC: Hepatocellular carcinoma; TCGA: The Cancer Genome Atlas; ML: Machine learning; NMR: Nuclear magnetic resonance; SHAP: SHapley Additive exPlanations; MRE: Magnetic resonance elastography; SVM: Support vector machine; UHN: University Health Network; ResNet50: Residual network 50-layer.

Table 8 Summary of major artificial intelligence applications in gastroenterology by clinical setting, evidence level, and implementation readiness.

Application	Clinical setting	Intended user	Evidence level	External validation	Regulatory status	Unresolved limitations
Colorectal polyp/adenoma detection (CADe)[61-65]	Screening/surveillance colonoscopy	Endoscopist	RCT meta-analysis	Yes	Approved in the United States, Europe, and Japan	No CRC-incidence/mortality data yet, slight increase in withdrawal time and non-neoplastic resections
Gastric neoplasia detection/characterization (CADe/CADx)[34,36-44]	Screening/surveillance endoscopy	Endoscopist	Prospective single/multicenter, including 2 single-center RCTs	Yes	Approved in Japan	Single-center RCTs, generalizability unclear
Colorectal polyp optical diagnosis (CADx, resect-and-discard)[66,67]	Screening colonoscopy	Endoscopist	Meta-analysis	Yes	Approved in Japan	Modest incremental benefit over unassisted diagnosis
IBD endoscopic/histologic activity assessment (PICaSSO/PHRI)[71-75]	IBD surveillance colonoscopy and pathology	Endoscopist, pathologist	Multicenter prospective	Yes	AI-assisted assessment of inflammatory activity in UC during colonoscopy approved in Europe and Japan	Not yet linked to hard outcomes (relapse, colectomy)
Esophageal cancer detection, staging and prognosis[20-24]	Screening and surveillance endoscopy	Endoscopist, including non-experts	Prospective multicenter, international multicenter CADe study	Yes	Approved in Japan	Lesion-margin delineation remains difficult, survival models validated only in geographically restricted cohorts
CE lesion detection[46-53,55]	Small bowel diagnostic work-up	Gastroenterologist	Mixed retrospective/prospective, meta-analysis	Yes	Approved in United States	Reading-time benefit not yet linked to patient outcomes
Barrett’s esophagus detection[21,23]	Screening and surveillance endoscopy	Endoscopist	Retrospective/model development and validation study	Yes	Approved in Europe	Limited explainability
Liver fibrosis staging, lesion characterization and hepatitis-subtype classification[107-113,116-119]	Hepatology outpatient/radiology	Hepatologist, radiologist	Retrospective, some multicenter	Yes	Not approved	Performance decline on external validation, some reclassifications lack an absolute ground truth
Pancreatic cancer detection (radiomics, EHR-based risk models)[82-85]	Radiology/population risk stratification	Radiologist	Multi-institutional retrospective plus prospective validation	Yes	Not approved	No survival/patient-outcome data yet, low pretest prevalence limits real-world PPV
Biliary/ERCP-related AI (stone prediction, stricture characterization, PEP prediction)[95-102,104]	ERCP/cholangioscopy	Endoscopist	Retrospective single/multicenter	Yes	Not approved	Low PPV for PEP prediction, models built largely on trial-derived data

AI: Artificial intelligence; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CE: Capsule endoscopy; CRC: Colorectal cancer; HER: Electronic health record; ERCP: Endoscopic retrograde cholangiopancreatography; IBD: Inflammatory bowel disease; PEP: Post-endoscopic retrograde cholangiopancreatography pancreatitis; PHRI: PICaSSO Histologic Remission Index; PICaSSO: Paddington International Virtual ChromoendoScopy Score; PPV: Positive predictive value; RCT: Randomized controlled trial; UC: Ulcerative colitis.

CURRENT PERSPECTIVE AND EMERGING PROSPECTS OF AI IN GASTROENTEROLOGY

Regulatory-approved systems

AI in gastroenterology is increasingly moving from experimental research toward regulatory-approved clinically applicable technologies. At present, most of the AI applications are in gastrointestinal endoscopy, particularly CADe and CADx. Regulatory approvals in the United States, Europe, and Japan include systems for real-time colorectal polyp detection, colorectal lesion characterization, gastric neoplasia detection, assessment of UC activity, and analysis of CE images. The approved AI systems and their respective clinical functions are summarized in Table 9.

Table 9 Artificial intelligence systems approved for clinical use in gastroenterology.

AI system name	Company	Clinical use	Regulatory authority ID
Japan - PMDA/MHLW
EndoBRAIN	Cybernet Systems Ltd.	CADx for differentiating neoplastic and non-neoplastic colorectal lesions using magnified endoscopy	PMDA: 23000BZX00372000
EndoBRAIN-EYE	Cybernet Systems Ltd.	CADe for detecting colorectal lesions during colonoscopy	PMDA: 30200BZX00208000
EndoBRAIN-UC	Cybernet Systems Ltd.	AI-assisted assessment of inflammatory activity in UC	PMDA: 30200BZX00136000
EndoBRAIN-Plus	Cybernet Systems Ltd.	CADx for predicting histological category of colorectal lesions	PMDA: 30200BZX00235000
CAD EYE (EW10-EC02)	Fujifilm Corporation	CADe/CADx for colorectal lesion detection and neoplastic/non-neoplastic differentiation	PMDA: 30200BZX00288000
CAD EYE (EW10-EG01)	Fujifilm Corporation	CADe for suspected esophageal squamous cell carcinoma and gastric neoplastic lesions	PMDA: 30400BZX00217000
EIRL Colon Polyp	LPIXEL Inc.	CADe for detecting colorectal polyps	PMDA: 30400BZX00259000
EndoBRAIN-X	Cybernet Systems Ltd.	CADx for differentiating colorectal neoplastic and non-neoplastic lesions using NBI	PMDA: 30500BZX00007000
gastroAI model-G2	AI Medical Service Co., Ltd.	CADe for detecting suspicious areas for early gastric cancer and adenoma	PMDA: 30600BZX00266000
United States-FDA
GI Genius™	Cosmo Artificial Intelligence - AI Ltd.	AI-assisted real-time detection of colorectal polyps and other mucosal lesions	FDA: K241887
EndoScreener	Chengdu Wision Medical Device Co., Ltd.	AI-assisted real-time detection of colorectal polyps	FDA: K211326
SKOUT® system	Iterative Health	AI-assisted real-time detection and localization of colorectal polyps	FDA: K253664
CAD EYE (EW10-EC02)	Fujifilm Corporation	AI-assisted detection of colorectal lesion candidates	FDA: K230751
NaviCam ProScan	Ankon Technologies Co., Ltd.	AI-assisted analysis of CE images for GI lesion detection	FDA: DEN230027
MAGENTIQ-COLO	Magentiq Eye Ltd.	CADe for real-time detection of colorectal lesions	FDA: K252178
CADDIE	Odin Medical Limited	AI-assisted detection of suspected colorectal polyps	FDA: K252586
Gastric Alimetry	Alimetry Ltd.	Analysis of gastric myoelectrical activity to aid diagnosis of gastric disorders	FDA: K252504
Deep Capsule®	Digestaid - Artificial Intelligence Development SA	AI-assisted analysis of CE videos for GI lesion detection	FDA: K250655
European Union - EUDAMED/EU MDR
CADDIE	Odin Medical Limited	AI-assisted detection of colorectal polyps	Basic UDI-DI: B-15065010840000
CADU	Odin Medical Limited	AI-assisted detection/characterization of dysplasia in Barrett’s esophagus	Basic UDI-DI: B-15065010840017
SMARTIBD	Odin Medical Limited	AI-assisted assessment of inflammatory activity in UC during colonoscopy	Basic UDI-DI: 5065010840OD04PF
GI Genius™ software	LINKVERSE S.R.L.	AI-assisted real-time detection of colorectal polyps	Basic UDI-DI: 8055060115WDCBB

AI: Artificial intelligence; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CE: Capsule endoscopy; EUDAMED: European Database on Medical Devices; EU MDR: European Union Medical Device Regulation; FDA: Food and Drug Administration; GI: Gastrointestinal; ID: Identification; MHLW: Ministry of Health, Labour and Welfare; NBI: Narrow-band imaging; PMDA: Pharmaceuticals and Medical Devices Agency; UC: Ulcerative colitis; UDI-DI: Unique device identifier-device identifier.

Emerging prospects

Beyond the organ-specific systems discussed above, several emerging AI technologies are still undergoing clinical validation that may significantly shape the future of gastroenterology. LLMs and generative AI are being explored for clinical reasoning, documentation, research, and extraction of information from unstructured EHRs. A systematic review of 21 early studies on applications of LLM models reported that GPT-4 demonstrated alignment with expert recommendations in pancreatic cyst management and high guideline compliance in postcolonoscopy management, while BioBERT showed promising performance in predicting drug-induced liver injury[127]. Another study demonstrated a practical application using a hybrid RegEx-LLaMA 2 pipeline to identify overt and recurrent gastrointestinal bleeding from clinical notes with high diagnostic performance[128].

Multimodal AI represents an emerging shift in gastrointestinal medicine from image-based analysis toward integrated, context-aware diagnostic systems. In CE, multimodal approaches combine video with inertial, magnetic, and physiological sensor data to improve anatomical localization and spatiotemporal interpretation[129]. Beyond endoscopy, integration of radiomics with genomic, molecular, and clinical data may improve prediction of lymph node metastasis of colorectal cancer and the recurrence of HCC after treatment[130].

Computer-assisted quality assessment is another emerging AI application that enables objective, real-time evaluation of procedural quality during gastrointestinal endoscopy by assessing mucosal visualization, illumination, image quality, and procedural duration. Computer-assisted quality systems are showing promise in improving adenoma detection, particularly for small and diminutive lesions[131].

LIMITATIONS AND CHALLENGES OF AI IN GASTROENTEROLOGY

Despite its growing applications, several challenges continue to limit the widespread adoption of AI in gastroenterology. A major limitation is the dependence of AI models on large, high-quality, and representative datasets. Inaccurate annotations, poor-quality images, missing clinical data, and underrepresentation of rare diseases can adversely affect model performance and introduce bias[132]. Limited generalizability is another concern. Many AI systems are developed using retrospective, single-center datasets and may perform poorly when applied to different populations, healthcare settings, or imaging platforms. Models are also susceptible to overfitting, dataset shift, and confounding factors, which can reduce their real-world applications[133]. The lack of prospective validation remains a barrier for clinical implementation. Although numerous studies report high diagnostic accuracy, robust prospective multicenter studies and RCTs demonstrating meaningful improvements in patient outcomes are still limited[133].

The practical deployment of AI in gastroenterology is constrained by fragmented healthcare data, which remain siloed across medical imaging, pathology, and EHR systems[133]. This fragmentation is complicated by the “black box” nature of DL, in which complex models and decision-making processes make it difficult for clinicians to understand the basis of individual predictions[132,134,135]. Explainable AI techniques, including saliency maps and Shapley additive explanations, may help address this limitation by providing interpretable information that enables clinicians to assess and verify AI-generated outputs[132,134,136]. Lack of interpretability may also contribute to automation bias and alert fatigue, potentially encouraging excessive reliance on AI recommendations without improving patient outcomes[132,133,136].

The use of AI introduces medico-legal and accountability challenges, particularly regarding the allocation of responsibility among clinicians, developers, and healthcare institutions when AI-assisted decisions contribute to diagnostic errors or patient harm[132,136]. Ethical and societal concerns include patient privacy, data security, cybersecurity risks, and the risk of widening healthcare disparities if AI systems are trained on non-representative datasets[135,136]. Federated learning has the potential to overcome privacy barriers by enabling AI models to be trained across decentralized sites without raw data transfer, thereby mitigating the data silo and security risks noted in the sources[133].

Finally, concerns regarding regulation, implementation, and cost-effectiveness remain significant obstacles. AI adoption is constrained by geographic, demographic, and healthcare-system differences, with variations in disease prevalence and clinical practice limiting model transferability across populations[132,136]. The predominance of research from high-income countries raises concerns that deploying Western-trained models in low- and middle-income settings without local validation may widen healthcare disparities[132]. Fragmented EHR, imaging, and pathology infrastructure, together with the need for ongoing technical support, data management, and staff training, can further hinder integration into existing clinical workflows[133-135]. Well-defined frameworks for oversight, accountability, and reimbursement are still evolving, while evidence supporting the long-term economic value of AI in gastroenterology remains limited[132]. Figure 5 summarizes the key limitations and challenges of AI in gastroenterology.

Figure 5 Key limitations and challenges of artificial intelligence in gastroenterology.

CONCLUSION

AI is rapidly transforming gastroenterology, with applications encompassing disease detection, risk stratification, endoscopic diagnosis, histopathological assessment, prognostication, and therapeutic decision-making. The translation of AI from research to clinical practice is already underway. The growing number of AI systems receiving regulatory authorization across the world reflects the increasing clinical maturity and integration of AI into gastroenterology. Despite many challenges, ongoing advances in AI and the integration of multimodal data, digital pathology, radiomics, genomics, and LLMs are expected to further improve precision gastroenterology. At present, AI should be viewed as a clinical decision-support tool, rather than a replacement for physician expertise. With further validation and responsible implementation, AI is likely to become an integral part of personalized, data-driven gastroenterology.