Artificial intelligence in gastroenterology practice: Current perspectives and emerging prospects

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.