Updates on the clinical use of artificial intelligence in upper gastrointestinal diseases: A comprehensive literature review

Table 1 Summary of literature for Barrett’s Esophagus

Disease	Ref.	AI architecture	Training/validation cohort	Targeted Endpoint	Key AI performance metrics	Performance compared to humans
BE	[8]	CNN-based CADe	Benchmarking image test set: 400 images; Benchmarking video test set: 188 videos	Detection of Barrett’s neoplasia	Images: Sensitivity 90%, Specificity 80%; videos: Sensitivity 91%, Specificity 82%	CADe system compared to general endoscopists — Images: Sensitivity 90% vs 74% videos: Sensitivity 91% vs 67%
	[9]	CNN-based CADe	Training: 75198 neoplastic images/videos (96 patients) and 1014973 non-neoplastic images/videos (65 patients); image validation: 471 images (34 patients); video validation: 75 videos (75 patients)	Detection of Barrett’s neoplasia	Sensitivity 93.8%, Specificity 90.7%, accuracy 92.0%	Non-expert endoscopists: Sensitivity 63.5%, specificity 77.9%, Accuracy 71.8% (P < 0.05 for all)
	[10]	CNN-based CADe	Pretraining: 494364 labeled endoscopic images; fine-tuning: 1704 esophageal high-resolution images (669 patients); benchmarking test set: 80 images (80 patients)	Detection of early neoplasia in BE	Sensitivity 93%, specificity 83%, accuracy 88%	53 general endoscopists: Sensitivity 72%, specificity 74%, accuracy 73%, all substantially inferior to the CADe system scores
	[11]	NLP	Development: 600 pathology reports (457 confirmed BE, 60 with dysplasia); validation: 400 reports (230 confirmed BE, 39 with dysplasia)	Identification of dysplasia in BE	Accuracy 98.7%, recall 92.3%, precision 100%	N/A
	[12]	NLP	Development/validation: 426 pathology notes (426 patients); full cohort application: 121808 individuals	Identification of intestinal metaplasia and dysplasia	Overall accuracy 97.6%-100%; BE: Sensitivity 98.9%, specificity 99.6%; gastric IM: Sensitivity 86.8%, specificity 99.0%	N/A
	[13]	CNN ensemble on digitized histology	542 patients (164 NDBE, 226 LGD, 152 HGD); 8596 training bounding boxes, 1946 validation, 840 test	Classification of BE dysplasia grade	Low grade dysplasia: Sensitivity 81.3%, specificity 100%; non-dysplastic BE and high grade dysplasia: Sensitivity and specificity > 90%	N/A
	[14]	Deep learning (CNN-based image and video analysis)	1970 pCLE videos; 897931 biopsy patches; 387 whole-slide images	Classification of BE dysplasia on pCLE	pCLE: Sensitivity for dysplasia 71%, overall accuracy 90%	N/A
	[16]	CNN (pretrained on ImageNet, fine-tuned) with object detection algorithm	916 dysplastic images (65 patients) and 919 non-dysplastic control images; 458 test images (225 dysplasia, 233 non-dysplasia)	Detection of early neoplasia in BE	Sensitivity 96.4%, specificity 94.2%, accuracy 95.4%	N/A
	[17]	CNN (indirectly supervised classifier + directly supervised segmentation/localization model)	Training: 148936 video frames (64 patients); Validation: 25161 images (11 patients); test: 264 i-scan images (44 patients); localization model: Trained on 94 delineated images (30 patients), tested on 86 images (28 patients)	Detection and localization of BE dysplasia	AUC 93%, Sensitivity 91%, specificity 79%; on endoscopist comparison subset: Sensitivity 96%, specificity 88%	6 non-expert endoscopists (same image subset): Sensitivity 79%, specificity 49%
	[18]	CNN (EfficientNetV2B2, pretrained, NBI endoscopic images)	724 cases total; test set: 160 NBI images from 86 patients with histological confirmation	Detection of BE on NBI	Sensitivity 94.3%; specificity 94.4%, accuracy 94.37%	N/A
	[19]	Vision Transformer (ViTBase16, pretrained)	1918 images from 267 patients (classified as dysplastic or non-dysplastic BE)	Classification of BE dysplasia	Sensitivity 97%, specificity 99%, accuracy 99%	N/A

AI: Artificial intelligence; CNN: Convolutional neural network; CADe: Computer-aided detection; BE: Barrett’s esophagus; NLP: Natural language processing; N/A: Not applicable; IM: Intestinal metaplasia; NDBE: Nondysplastic BE; LGD: Low-grade dysplasia; HGD: High-grade dysplasia; pCLE: Probe-based confocal laser endosmicroscopy; AUC: Area under the curve; NBI: Narrow-band imaging.

Table 2 Summary of literature for gastroesophageal reflux disease

Disease	Ref.	AI architecture	Training/validation cohort	Targeted endpoint	Key AI performance metrics	Performance compared to humans
GERD	[21]	ANN with TWIST evolutionary optimization (symptom questionnaire-based)	577 GERD patients and 94 healthy controls; QUID 41-item questionnaire data	Diagnosis of GERD from symptoms	Sensitivity 99.1%, specificity 99.8%, accuracy 99.2%	N/A
	[22]	ANN, decision trees, and logistic regression (data mining on 15-item symptom questionnaire)	132 patients with upper GI symptoms; 15-item diagnostic questionnaire	Diagnosis of GERD from symptoms	Sensitivity 70%-75%, specificity 63%-78%, AUC 0.787	N/A
	[23]	Large language model (ChatGPT, version March 14, 2023)	23 prompts × 3 submissions = 69 total queries; rated by 3 gastroenterologists and 8 patients	GERD management recommendations	91.3% appropriate responses (29.0% completely, 62.3% mostly); 78.3% contained specific guidance; 70% of prompts yielded inconsistent appropriateness across repeats	N/A
	[25]	Deep learning (DenseNet-121) with explainable AI (XAI)	2081 endoscopic images	LA-grade classification of GERD	AUC 0.968, classification accuracy 86.7%	Junior endoscopists: Accuracy 71.5%; experienced endoscopists: Accuracy 77.4% (DL model significantly higher)
	[26]	DL + CNN (GERD-VGGNet; VGG16 pretrained on ImageNet with transfer learning)	Development set: 464 cases (247 LA A-B, 155 LA C-D, 62 normal in conventional mode; augmented to 603000 training images); Test set: 32 images	LA-grade classification of GERD	Sensitivity 99.8%, specificity 99.7%, overall accuracy 98.9%	NBI test set — trainee 1 overall accuracy 75.0%; trainee 2 overall accuracy 65.6% (AI vs trainee 2 significant at 95% confidence)
	[27]	Two-stage: DL (Gerd-VGGNet/VGG16) for feature extraction + machine learning (Random Forest) classifier (GerdNet-RF)	Development set: 671 NBI images (244 LA A-B, 229 LA C-D, 198 normal); Test set: 32 images (12 LA A-B, 10 LA C-D, 10 normal)	LA-grade classification of GERD	Test set accuracy 92.5%, validation set accuracy 98.4%	N/A
	[28]	Prototype AI software for AIBI from pH-impedance tracings	325 subjects (117 off therapy, 93 on therapy, 115 asymptomatic controls)	Prediction of GERD symptom improvement	AIBI ratio AUC 0.661	N/A
	[29]	AI/machine learning with deep residual network (ResNet) for image recognition of impedance-pH tracings	106 patients; 7939 impedance events	Automated reflux episode and PSPW detection	Reflux episode accuracy 87%; PSPW accuracy 82%	N/A

ANN: Artificial Neural Network; AI: Artificial intelligence; GERD: Gastroesophageal reflux disease; N/A: Not applicable; AUC: Area under the curve; DL: Deep learning; CNN: Convolutional neural network; LA: Los Angeles; NBI: Narrow-band imaging; AIBI: Automated AI extraction of baseline impedance; PSPW: Postreflux swallow-induced peristaltic wave.

Table 3 Summary of literature for gastroesophageal varices

Disease	Ref.	AI Architecture	Training/validation cohort	Targeted endpoint	Key AI performance metrics	Performance compared to humans
Varices	[33]	Deep CNN (ENDOANGEL, real-time endoscopic AI system)	8566 variceal images (3021 patients) + 6152 normal images (3168 patients)	Detection and characterization of GEV	EV detection accuracy 97.0%; GV detection accuracy 92.0%	Endoscopists: EV detection 93.94% (P < 0.01); GV detection 84.43% (P < 0.05)
	[34]	Explainable AI platform (ENDOANGEL-GEV; multiple deep learning models for segmentation, RC classification, and variceal grading)	Testing dataset: 1351 EV + 811 GV varices; validation dataset: 2402 EV + 12383 GV varices; prospective dataset: 161 patients	GEV detection, grading, and risk stratification	EV detection sensitivity 90.79% (validation); RC classification EV accuracy 94.62%; EV grading accuracy 94.57%; risk stratification EV accuracy 97.69% (prospective)	Prospective comparison — endoscopists: RC classification EV 66.92% (P < 0.001); EV grading 75.97% (P < 0.001)
	[35]	Deep learning — EfficientNet V2 (best of 6 CNN/transformer-based models; transfer learning from ImageNet; Grad-CAM visualization)	Training: 675 images (439 control, 236 bleeding) augmented to 2000; validation: 400 images; external test: 400 images (200 control, 200 bleeding)	12-month EV bleeding risk prediction	Validation: Accuracy 0.910, F1 0.905; external test: Accuracy 0.893, precision 0.911, F1 0.890	Endoscopist #1: Accuracy 0.800; endoscopist #2: Accuracy 0.763; With AI assistance: Accuracy improved to 0.938 (+17.3%) and 0.908 (+19.0%), respectively
	[36]	Multimodal machine learning: DL image models (EfficientNet best; CNN + transformer-based) combined with clinical structured data via stacking ensemble (H2O AutoML; DL, XGBoost, RF, GBM, GLM)	810 EV images augmented to 2400 (2000 training, 400 validation); 341 patients with clinical data; external test dataset from second center	12-month EV bleeding prediction	Test dataset (stacking model): Accuracy 0.932, sensitivity 0.952, F1 0.879; AUC 0.975	N/A
	[37]	Machine learning (Random Forest algorithm) using clinical variables (EVendo score)	Training: 238 patients; validation: 109 patients (both from 3 hospitals)	Prediction of EV and VNT	EV: AUROC 0.84 (training), 0.82 (validation); VNT: AUROC 0.74 (training), 0.75 (validation)	N/A
	[38]	Machine learning-derived clinical score (EVendo score; external validation)	External validation cohort: 103 patients (66 with GEV, 12 with VNT)	Prediction of GEV and VNT	AUROC for GEV 0.75; AUROC for VNT 0.59; EVendo ≤ 3.90: NPV 94% for VNT, spared 20.4% of EGDs, missed VNT 19%	N/A
	[39]	Deep learning (DeepLabV3+ CNN) for automated CT liver/spleen segmentation and volumetry; spleen volume-to-platelet ratio index	DL segmentation algorithm trained on 813 labeled CT datasets; clinical study: Derivation 239 patients, validation 180 patients; Follow-up endoscopy: 219 patients	Detection of high-risk esophageal varices	Derivation AUC 0.835; validation cutoff > 3.78: Sensitivity 69.4%, specificity 78.5%; cutoff > 1.63: 100% sensitivity	N/A
	[40]	Machine learning-based CT RM using portal venous-phase enhanced CT (liver, spleen, esophagus ROIs)	Training/internal validation: 458 patients (218 mild EV, 240 HREV); external validation: 499 patients (159 mild EV, 340 HREV)	Diagnosis of high-risk esophageal varices	HREV RM — internal validation: AUC 0.834, sensitivity 0.977, specificity 0.969; external validation: AUC 0.736, sensitivity 0.690, specificity 0.762	N/A
	[41]	Machine learning (Random Forest, AdaBoost, SVM, linear regression) using CT morphology + serum biomarkers + clinical features	224 patients (5-fold cross-validation)	Prediction of esophageal variceal bleeding	Test set — random forest AUC 0.818; AdaBoost AUC 0.804; linear regression AUC 0.763	N/A

RM: Radiomics model; AI: Artificial intelligence; CNN: Convolutional neural network; GEV: Gastroesophageal varices; EV: Esophageal variceal; GV: Gastric varices; RC: Red color; AUC: Area under the curve; VNT: Varices needing treatment; AUROC: Areas under the receiver-operating-characteristics curves; NPV: Negative predictive value; EGD: Esophagogastroduodenoscopy; DL: Deep learning; CT: Computed tomography; ROIs: Regions of interest; HREV: High-risk oesophageal varices.

Table 4 Summary of literature for esophageal cancer

Disease	Ref.	AI Architecture	Training/validation cohort	Targeted endpoint	Key AI performance metrics	Performance compared to humans
Esophageal cancer	[47]	Deep Neural Network-based Computer-Aided Detection (DNN-CADe)	Training: 2428 images (1332 abnormal, 1096 normal) from 746 patients; validation: 187 images from 52 patients	Detection of early esophageal SCC	AUC > 96%; sensitivity 97.8%, specificity 85.4%, accuracy 91.4%	Endoscopists alone: Sensitivity 74.2%, accuracy 81.7%, NPV 79.3%; With DNN-CADe assistance: Sensitivity 89.2%, accuracy 91.1%, NPV 90.4% (all significant)
	[48]	Deep learning (ENDOANGEL-ELD; multiple CNN modules for esophageal lesion detection)	191744 esophagogastroscopic images from multiple institutions (training); 3117 patients enrolled in RCT (1556 experimental, 1561 control)	Detection of high-risk esophageal lesions	System: Sensitivity 89.7%, specificity 98.5%, accuracy 98.2%	HrEL detection rate: AI-assisted 1.8% vs unassisted 0.9% (P = 0.029); early esophageal cancer: 0.7% vs 0.2% (P = 0.031)
	[49]	Deep learning (iCLE system; Inception-ResNet V2 with image recognition, video diagnosis, and quality judgment modules) on pCLE	Training/validation: 25056 images (2803 patients); testing: 2442 images (226 patients); prospective video: 187 videos (67 lesions, 54 patients)	Diagnosis of esophageal squamous neoplasm	Image test: Accuracy 98.3%, sensitivity 95.3%, specificity 98.8%; video: Accuracy 90.9%, sensitivity 92.0%, specificity 90.2%	Expert endoscopists: Sensitivity 80.4%, NPV 87.7% (iCLE superior: P < 0.001, P = 0.003); non-experts with iCLE assistance: Accuracy improved 83.6% to 88.6%, sensitivity 76.0% to 89.8%
	[50]	Deep learning (CNN-based AI system) on EUS images	16110 EUS images from 631 cases (5 SEL categories); 4:1 development-to-test ratio	Classification of upper GI SELs	5-class accuracy 86.1%; GIST vs non-GIST: Sensitivity 98.8%, specificity 67.6%, accuracy 89.3%	AI 5-class accuracy significantly higher than all endoscopists; GIST sensitivity and accuracy significantly higher than all endoscopists (both experts and non-experts)
	[51]	CNN-based real-time detection and segmentation (YOLOv8s-seg + MobileNetv2) on EUS images	1445 esophageal SEL cases	SEL detection and layer identification	Detection: Precision 92.2%, F1 81.9%; layer 2/3 vs 4 differentiation accuracy 76.5%	Senior endoscopists: 74.9%-79.8% (P = 0.585, comparable); Junior endoscopists: 65.6%-66.7% (P = 0.045, AI significantly higher)

AI: Artificial intelligence; DNN: Deep neural network; CADe: Computer-aided detection; SCC: Squamous cell cancer; AUC: Area under the curve; NPV: Negative predictive value; CNN: Convolutional neural network; HrEL: High-risk esophageal lesion; RCT: Randomised controlled trial; iCLE: Intelligent confocal laser endomicroscopy; EUS: Endoscopic ultrasound; SEL: Subepithelial lesion; GI: Gastrointestinal; GIST: Gastrointestinal stromal tumor.

Table 5 Summary of literature for gastric cancer

Disease	Ref.	AI architecture	Training/validation cohort	Targeted endpoint	Key AI performance metrics	Performance compared to humans
Gastric cancer	[46]	Deep learning (ENDOANGEL-LD; CNN-based real-time lesion detection during white light endoscopy)	907 patients AI-first group; 905 patients routine-first group	Gastric neoplasm miss rate reduction	AI-first miss rate 61% vs routine-first 27.3% (relative risk 0.224, P = 0.015)	Routine (unassisted) endoscopy: Gastric neoplasm miss rate 273%; AI-assisted: 6.1% (P = 0.015; 77.6% relative reduction)
	[58]	Multimodal deep learning — incremental Supervised Contrastive Learning Model (iSCLM): ResNet-34 (CT features) + ResNet-18 with Graph Attention Network (pathology WSI features)	Development: 1208 patients; incremental: 795 patients; external test 1: 162 patients; external test 2: 222 patients; prospective: 132 patients	Prediction of neoadjuvant chemotherapy response	Prospective cohort: AUROC 0.846; external test 1: AUROC 0.866; external test 2: AUROC 0.876	N/A
	[59]	Deep learning (GRAPE; 3D noncontrast CT with pixel-level segmentation, morphometric and contextual radiological feature integration)	Development: 3470 GC + 3250 non-GC cases; internal validation: 1298 cases; external validation: 18160 cases; opportunistic screening: 78593 CT scans	Gastric cancer identification and screening	Internal validation AUC 0.970; External validation AUC 0.927	GRAPE outperformed radiologists: Sensitivity improved by 21.8%, specificity by 14.0%
	[60]	AI-driven biomarker discovery (ESGCmiRD; machine learning integrating miRNA expression patterns, network analysis, and regulatory modeling)	miRNA expression data from 10071 patients (GEO + TCGA databases); validated on GSE211692, TCGA-STAD, and institutional cohort	Early-stage GC detection	5-miRNA signature AUC: 0.986 (test set)	N/A
	[61]	CNN-based whole-slide image classification algorithm	Training: 2434 whole-slide images; prospective validation: 7440 biopsy specimens; pathologist impact study: 150 cases assessed by 6 pathologists	Classification of gastric biopsy pathology	AUROC 0.9790 (NFD vs positive); Epithelial tumors: Sensitivity 100%, specificity 97.49%; 52% reduction in review time	N/A

AUROC: Areas under the receiver-operating-characteristics curves; GEO: Gene Expression Omnibus; TCGA: Cancer Genome Atlas; AI: Artificial intelligence; AUC: Area under the curve; CNN: Convolutional neural network; CT: Computed tomography; GC: Gastric cancer; GRAPE: Gastric Cancer Risk Assessment Procedure with Artificial Intelligence; N/A: Not applicable; NFD: Negative for dysplasia.