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

Abstract

Artificial intelligence (AI) is increasingly being integrated into gastroenterology, with growing applications in upper gastrointestinal (GI) disease detection, diagnosis, risk stratification, and treatment guidance. We performed a comprehensive literature review of the clinical use of AI in upper GI pathologies, focusing on Barrett’s esophagus, gastroesophageal reflux disease, esophagogastric varices, esophageal cancer, and gastric cancer. A systematic search of PubMed/ MEDLINE, EMBASE, Google Scholar, and the Cochrane Library was conducted from October 2024 through March 2026 using Medical Subject Headings terms and free-text keywords related to AI and upper GI disease. Eligible English-language studies included original research, reviews, systematic reviews, and meta-analyses addressing diagnostic, prognostic, or therapeutic applications of AI. Across disease states, AI demonstrated strong performance in image recognition, endoscopic lesion detection, histopathologic interpretation, bleeding risk prediction, and automated clinical decision support. In several studies, AI systems matched or outperformed general endoscopists in detecting Barrett’s-related dysplasia, grading reflux esophagitis, identifying and risk-stratifying varices, and recognizing early esophageal and gastric neoplasia. Emerging multimodal and radiomics-based models also showed promise for noninvasive screening and treatment response prediction. Despite encouraging results, most studies remain retrospective and heterogeneous. Prospective multicenter validation, regulatory oversight, and ethical safeguards are needed before widespread clinical implementation.

Key Words: Artificial intelligence; Barrett’s esophagus; Gastroesophageal reflux disease; Gastroesophageal varices; Esophageal cancer; Gastric cancer

Core Tip: Artificial intelligence (AI) is reshaping the management of upper gastrointestinal diseases by improving lesion detection, diagnosis, risk stratification, and clinical decision-making. AI has demonstrated diagnostic accuracy comparable to that of general endoscopists in several upper gastrointestinal conditions. While these findings are promising, prospective validation and standardized implementation are needed before widespread clinical adoption.

INTRODUCTION

Artificial intelligence (AI) is rapidly evolving with a massive potential to transform the medical field. Clinical applications of AI are being explored for most clinical specialties, offering novel tools for disease prediction, diagnosis, risk stratification and personalized patient care[1]. AI-based clinical decision support systems have emerged, supporting clinicians with complex decision-making processes. Published literature discusses uses of AI across several fields in medicine such as radiology, radiation oncology, oncology and surgery[2-5].

In gastroenterology, AI systems can leverage massive amounts of data to build predictive models and have potential to reshape the field and revolutionize patient care. Some studies explored AI applications in the endoscopic detection and diagnosis of malignant gastrointestinal (GI) lesions, histopathology analysis of biopsy specimens, prediction of treatment response in patients with inflammatory bowel disease and liver cirrhosis[6]. AI and machine learning (ML) in hepatology also has the potential to support and guide complex liver transplant decisions[7].

This manuscript is an updated comprehensive literature review focusing on the clinical uses of AI in upper GI pathologies, covering Barrett’s esophagus (BE), gastroesophageal reflux disease, esophagogastric varices, esophageal cancer and gastric cancer (GC).

A comprehensive literature search was conducted between October 2024 and March 2026 to identify relevant studies evaluating the application of AI in upper GI pathologies. Electronic databases including PubMed/MEDLINE, EMBASE, Google Scholar, and the Cochrane Library were systematically searched for eligible publications.

The search strategy incorporated a combination of Medical Subject Headings terms and free-text keywords. The following terms and their Boolean combinations were used: “artificial intelligence”, “machine learning”, “deep learning”, “clinical decision support systems”, “upper gastrointestinal diseases”, “Barrett’s esophagus”, “gastroesophageal reflux disease”, “esophagogastric varices”, “esophageal cancer”, “gastric cancer”, and “medico-legal considerations”. Reference lists of included articles were also manually screened to identify additional relevant studies.

Eligible studies included original research articles, case reports, narrative reviews, systematic reviews, and meta-analyses that addressed the use of AI or ML in the diagnosis, management, or prognostication of upper GI conditions. Both qualitative and quantitative studies were considered. Articles were limited to those published in English. Conference abstracts without full-text availability, editorials, and studies lacking sufficient methodological detail were excluded.

Titles and abstracts were screened for relevance, followed by full-text review of selected articles. Data extracted from each study included study design, population characteristics, type of AI model or intervention, clinical application (diagnosis, risk stratification, or treatment guidance), and key outcomes. Given the heterogeneity in study designs, AI methodologies, and reported outcomes, a narrative synthesis approach was employed to summarize findings. Key themes were identified, including diagnostic accuracy, clinical integration, and medico-legal implications of AI in upper GI practice.

DISCUSSION

BE

BE is a precursor to esophageal adenocarcinoma (AC). It is traditionally diagnosed by endoscopic visualization of columnar-lined esophagus extending above the gastroesophageal junction and confirmed via biopsy showing intestinal metaplasia with goblet cells. Advancements have been made over the past few years utilizing AI processing for aid in diagnosis of many conditions including Barrett’s esophagus.

There are four main categories of AI that aid in the diagnosis of BE. These categories are computer-aided systems, natural language processing (NLP)/text-based systems, deep learning (DL) on histology and biopsy images, and real-time and/or video analysis. A summary of significant studies in each category is included here.

Computer-aided detection (CADe) systems assist in the detection of suspicious areas in images. These systems help estimate the likelihood of disease and can act as a decision-support tool. One study evaluated a CADe system when evaluating dysplastic and non-dysplastic BE images and videos and when comparing the CADe to general endoscopists in identifying subtle Barrett’s neoplastic lesions on video test, the sensitivity of the CADe system was 91% vs the general endoscopist’s sensitivity of 67%[8]. Another study comparing a CADe system to 6 general endoscopists was conducted and found that CADe detected Barrett’s neoplasia with a sensitivity of 93.8% and specificity of 90.7%, overall higher compared with the endoscopists’ performance of 63.5% and 77.9%, respectively (P < 0.05)[9]. In another study on the detection of early neoplasia in BE, CADe system vs those of the general endoscopists were 88% vs 73% accuracy, 93% vs 72% sensitivity, and 83% vs 74% specificity[10].

NLP and text-based systems use AI to understand, interpret, and analyze language and can aid in the identification of dysplasia in BE histology reports. In one study, a set of 600 histology results, some with BE with dysplasia and some without dysplasia, were compared with manual review. An NLP algorithm was utilized, and it identified dysplasia with 98.7% accuracy, 92.3% recall, and 100.0% precision[11]. In another study, 121000 esophagogastroduodenoscopy (EGD) and 426 pathology notes were used to validate a NLP algorithm to identify intestinal metaplasia, dysplasia, and tumors. The accuracy of the NLP algorithm was 97.6% to 100%. For the Barrett’s esophagus group, 98.9% sensitivity and 99.6% specificity were reported[12].

Convolutional neural networks (CNNs) are a type of DL model that can classify images and aid in object detection. In the context of BE, it can aid in the detection of malignancy in biopsy samples. In one study, after histology slides with BE dysplasia were identified by expert pathologists, a DL model was trained using these images. Sensitivity and specificity for low-grade dysplasia was 81.3% and 100%, respectively, and > 90% for non-dysplastic BE and high grade dysplasia[13]. In probe-based confocal laser endosmicroscopy (pCLE) for in vivo histological observation, utility can be limited by low sensitivity. However, when DL-based analysis was used to improve diagnostic accuracy in identifying BE dysplasia, the model achieved a high sensitivity for dysplasia (71%) and an overall accuracy of 90%[14]. When DL is used to identify the most critical 2D image sections within large 3D pathology data sets, it can reduce the burden on pathologists and improve detection sensitivity of malignancy in BE biopsies[15].

Real-time and video analysis technologies, some based on CNN, can aid in the detection of BE and early esophageal neoplasia. Three studies utilized CNN-based object detection algorithms for BE dysplasia detection. The sensitivities ranged from 91%-96.4% while the specificities ranged from 79%-94.4%[16-18]. Another study used a vision transformer (ViT) model derived from natural language processing to aid in the classification of early esophageal neoplasia on video. Narrow band imaging ViT had a reported sensitivity of 97% and specificity of 99%, even higher than CNN based models[19].

AI demonstrates high diagnostic accuracy for detecting Barrett’s esophagus, with performance often exceeding that of general endoscopists. To confirm the clinical utility of AI in diagnosing BE, prospective randomized controlled trials evaluating real-time clinical performance and patient outcomes are needed (Table 1).

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.

Gastroesophageal reflux disease

Gastroesophageal reflux disease (GERD) is a chronic condition in which stomach acid flows upwards into the esophagus. Symptoms can include chest pain, dysphagia, and/or regurgitation. There are three recognized categories: Non-erosive reflux disease, erosive esophagitis, and disease-causing BE. There is no single gold-standard diagnostic test for GERD, the diagnosis relies on typical symptoms, endoscopy, reflux monitoring, and symptom improvement with proton pump inhibitor therapy[20].

AI demonstrates excellent performance when applied to GERD symptom-based questionnaires. One study utilized the TWIST system, a ML methodology designed to optimize the performance of artificial neural networks (ANN), to interpret results from a GERD questionnaire. The system was found to have a specificity, sensitivity, and accuracy of 99.8%, 99.1%, and 99.2%, respectively[21]. Another study designed a 15-item questionnaire to diagnose GERD based on clinical symptoms. After patients completed the questionnaire and underwent endoscopy, the results of the questionnaire were interpreted utilizing ANN. The sensitivity and specificity of the new symptom score were 70%-75% and 63%-78%, respectively[22]. These validated systems can be used to clinically diagnose GERD with high accuracy.

Large language model (LLM)-linked chatbots may be an efficient source of clinical recommendations for healthcare providers and patients. In one study, ChatGPT provided largely appropriate responses (91.3%), although some inappropriateness (8.7%) and inconsistency were reported[23]. Another study queried 5 LLM-linked chatbots using patient cases, and responses were graded according to whether the recommendations aligned with the Society of American Gastrointestinal and Endoscopic Surgeons guidelines. Surgeons were given accurate recommendations for the surgical management of GERD in an adult patient between 42.9%-85.7% of queries[24]. These results showcase some promising results as well as some pitfalls in the use of LLM chatbots to provide recommendations for the treatment of GERD.

The Los Angeles classification can be used to quantify GERD disease severity during endoscopy. The use of AI to aid in the endoscopic identification and classification of GERD, specifically using DL-based models, has been explored in multiple studies. One study evaluated the performance of a DL-based model in the classification of GERD compared to junior and experienced endoscopists. The reported classification accuracy of 86.7% was significantly higher than the junior (71.5%) and experienced (77.4%) endoscopists[25]. Other studies have proposed and tested novel models using a combination of DL, CNN, and ML. Overall accuracy for one such model utilizing DL and CNN was reported to be 98.9% with sensitivity of 99.8% and specificity of 99.7%[26]. Another study proposed and tested a two-stage model utilizing DL and ML and reported an accuracy of 92.5% ± 2.1%[27].

pH-impedance studies can diagnose GERD and assess therapeutic efficacy via the detection of acidity and impedance/electrical signals. The test involves the insertion of a catheter into the esophagus to collect this information. Manual analysis of pH-impedance tracings is time-consuming, so its use is limited. AI has the potential to streamline the interpretation of pH-impedance studies. Novel AI metrics show potential in predicting GERD treatment outcome. In one study, a prototype AI software was compared to manual extraction of data and AI outperformed manual extraction (area under the curve 0.661 vs 0.313)[28]. Another study on AI in pH-impedance interpretation showed that the accuracy of AI identification for reflux episodes and post reflux swallow-induced peristaltic wave index achieved 87% and 82%[29].

Overall, AI has demonstrated high diagnostic performance in GERD diagnosis and treatment. Applications span from questionnaire-based diagnostic models to endoscopic assessment and automated analysis of pH-impedance studies (Table 2).

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.

Gastroesophageal varices

Gastroesophageal varices (GEV) develop as a result of portal hypertension (PH), which develops as a consequence of increased resistance to portal blood flow. The increased resistance is due to a combination of structural changes, such as liver fibrosis, nodules or vascular occlusion, and dynamic changes, such as increased release of vasoconstrictors[30]. Esophageal variceal (EV) bleeding is a fatal complication and is the most common cause of upper GI bleeding[31]. EV develop in up to 40% of patients with compensated cirrhosis and up to 85% of patients with decompensated cirrhosis[10]. Variceal bleeding occurs at a rate of up to 15% per year, and the 6-week mortality is between 15% and 25%[32]. EV bleeding also tends to have high recurrence rates of 60% within 1-2 years without treatment[33].

Hepatic venous pressure gradient measurement is the diagnostic gold standard for PH, and EGD is the diagnostic gold standard for EV[30], but it is invasive, resource-intensive, and subject to interobserver variability in grading variceal size and red color (RC) signs. These limitations have made GEV an attractive target for AI, particularly for automated endoscopic detection, objective risk stratification, and development of noninvasive triage tools that may reduce unnecessary screening endoscopy.

One of the earliest and most clinically relevant applications of AI in this area has been computer-aided endoscopic assessment. In a multicenter retrospective study, Chen et al[33] developed and validated ENDOANGEL, a real-time deep CNN for EGD-based assessment of GEV. Using thousands of endoscopic images, the system achieved very high diagnostic performance for variceal detection, often surpassing endoscopists in the study. Overall, it achieved a higher detection accuracy for EV (97% vs 93.94%, P < 0.01) and gastric varices (GV) (92.00% vs 84.43%, P < 0.05) compared to the endoscopists. It also achieved a higher detection rate for RC signs of EV (84.21% vs 73.45%, P < 0.01) and red spots of GV (85.26% vs 77.52%, P < 0.05). Notably, the mean analysis time was 0.13 seconds per image for the AI system vs 18.75 seconds for endoscopists, suggesting a potential role for real-time decision support during endoscopy. Beyond that, Wang et al[34] developed ENDOANGEL-GEV, an explainable AI platform composed of multiple models designed to segment varices, classify variceal grade, and identify RC signs in real time. In the validation dataset, the system achieved a sensitivity of 90.79% for detecting EV, a sensitivity of 99.79% and specificity of 92.54% for predicting esophageal RC signs, and accuracies above 93% for classifying variceal grades. In a prospective comparison with endoscopists, ENDOANGEL-GEV significantly outperformed endoscopists in classification of RC signs in EV (94.62% vs 66.92%, P < 0.001) and in GV (94.92% vs 69.49% P < 0.001). It also showed a higher accuracy in grading both EV and GV, and an overall better risk assessment. The authors estimated that use of the system could help 12.31% more patients receive the correct intervention, highlighting the potential of AI not only to detect varices but also to standardize management-relevant endoscopic interpretation.

AI has also been applied to predict future variceal bleeding from endoscopic images alone. In a multicenter retrospective study, Hong et al[35] trained six DL models to classify the 12-month bleeding risk of EV using endoscopic images. EfficientNet had the best performance, achieving an accuracy of 0.910 in the validation dataset and 0.893 in the external test dataset, outperforming two endoscopists whose accuracies were 0.800 and 0.763. Importantly, when endoscopists were assisted by AI, their accuracies improved by 17.3% and 19.0%. A further step has been the use of multimodal ML (MMML), which combines endoscopic images with structured clinical data. Wang et al evaluated an automated MMML framework integrating endoscopic image features with laboratory and clinical variables to predict 12-month EV bleeding. After the development of several DL image models, the best-performing stacking model achieved an accuracy of 0.932, sensitivity of 0.952, and F1-score of 0.879 in the test dataset, outperforming conventional indices such as MELD, Child-Pugh, platelet ratio index, and Fibrosis-4 score[36]. This is clinically important because bleeding risk is not determined by endoscopic appearance alone; multimodal AI may better reflect the complex interaction between PH severity, liver dysfunction, and endoscopic morphology.

In parallel, AI has been used to create noninvasive tools to identify patients unlikely to harbor varices needing treatment (VNT), with the goal of reducing unnecessary screening EGDs. A notable example is the EVendo score, a ML-derived model based on readily available clinical variables including international normalized ratio, aspartate transaminase, platelet count, blood urea nitrogen, hemoglobin, and ascites. In its original development and validation study, EVendo achieved areas under the receiver-operating-characteristics curves of 0.84 and 0.82 for predicting EV, and 0.74 and 0.75 for predicting VNT[37]. Using a cutoff below 3.90, the score could have spared 30.5% of EGDs while missing only 2.8% of VNT. External validation showed that an EVendo score of 3.90 or less had a negative predictive value of 94% for VNT, and spared more endoscopies than Baveno VI criteria, although with somewhat more missed high-risk cases[38].

Another rapidly expanding area in noninvasive variceal risk assessment is AI-based radiology and radiomics, where computed tomography (CT)-derived DL and radiomic models (RM) attempt to infer PH severity and variceal risk without EGD. Lee et al[39] used a DL algorithm to automatically measure spleen and liver volumes on CT and developed a spleen volume-to-platelet ratio index for high-risk varices in compensated B-viral cirrhosis. In the validation cohort, a cutoff greater than 3.78 yielded a sensitivity of 69.4% and specificity of 78.5%, while a lower cutoff greater than 1.63 detected all high-risk varices. In another study, ML-based CT RM could diagnose high-risk EV with a higher accuracy compared to the expanded Baveno VI criteria[40]. More recently, Zhao et al[41] developed ML models combining CT morphology, serum biomarkers, and basic clinical characteristics to predict variceal bleeding; in their test dataset, random forest achieved an area under the curve (AUC) of 0.818 and adaptive boosting an AUC of 0.804. Together, these studies suggest that AI-enhanced imaging may become an important noninvasive complement to EGD, particularly for screening and pre-endoscopic risk stratification.

Recent systematic reviews support the promise of these approaches but also underscore important limitations. A 2024 systematic review of AI applications in EV found that models based on endoscopy, CT imaging, laboratory variables, and clinical scores often performed comparably to or better than traditional assessment, but the evidence base remains heterogeneous[42]. Another 2024 methodological systematic review concluded that machine-learning models show significant promise for predicting bleeding risk and grading varices, yet many studies carry unclear or high risk of bias because of limited sample sizes, incomplete reporting, and inadequate external validation[43].

Overall, AI use in GEV is an emerging field showing promise in terms of endoscopic detection and characterization of varices, prediction of risk of variceal bleeding using multimodal data, and noninvasive screening strategies to identify patients who may safely defer endoscopy. However, before these technologies can be widely adopted, larger prospective multicenter trials, standardized outcome definitions, and robust external validation across different populations and endoscopy platforms are still needed. Those technologies are best viewed as adjuncts to expert clinical judgement rather than replacement for guideline-based endoscopic evaluation (Table 3).

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.

Esophageal cancer

Esophageal cancer is the seventh most common cancer and sixth leading cause of cancer-related death worldwide[44]. There are two major causes of esophageal cancer: Squamous cell cancer (SCC) and AC. SCC is the most common subtype in the world accounting for 90% of the cases; however, AC is the most common type of esophageal cancer in the United States with increasing incidence. AC is typically caused as a sequela of chronic GERD, where stratified squamous epithelium undergoes metaplasia to columnar epithelium resulting in Barrett’s esophagus. SCC is associated with pathological and molecular changes that occur with lifestyle habits such as extensive alcohol use, tobacco use, consumption of hot beverages, and intake of food rich in nitrosamine compounds that lead to genetic mutations[44,45]. SCC is graded based on level of keratinization, nuclear pleomorphism, and mitotic activity where well-differentiated tumors demonstrate increased keratinization with low mitotic activity[46].

Endoscopy is an important modality to detect esophageal cancer. DL has been studied extensively in its use for lesion detection, especially in the early precancerous phase. One study utilized a novel system of CADe with deep neural network (DNN) to localize early esophageal squamous cell carcinoma. This DNN-CADe system had a diagnostic sensitivity of 97.8%; therefore, the diagnostic ability of endoscopists with the use of DNN-CADe improved in terms of sensitivity (74.2% vs 89.2%), accuracy (81.7% vs 91.1%), and NPV (79.3% vs 90.4%)[47]. Additionally, ENDOANGEL-ELD (esophageal lesion detection) system consisted of two major functions and five CNN modules in a prospective clinical trial utilizing 191744 esophagogastroscopic images from multiple medical institutions in China where 3117 patients were recruited to be evaluated in the experimental group with CNN assisted endoscopy or the unassisted endoscopy control group. In this study by Li et al[48], the intention-to-treat population has a statistically significant detection rate of high-risk esophageal lesions (HrELs) including esophageal cancer and precancerous lesions, as compared to the control group (1.8% vs 0.9%, P = 0.029). This also holds true in the per-protocol analysis for HrEL detection rate in the thoracic esophagus (1.6% vs 0.7%, P = 0.018) and in the detection of early esophageal cancer (0.7% vs 0.2%, P = 0.031).

Early detection of esophageal neoplasms is being studied with the use of optical imaging such as the use of pCLE for in vivo histological observation. The intelligent confocal laser endomicroscopy (iCLE) system was developed utilizing images and videos for image recognition training such that it achieved an accuracy of 98.3%, sensitivity of 95.3%, and specificity of 98.8% when evaluating neoplasms. Even when compared to expert endoscopists with extensive experience, iCLE system has both a statistically significant improvement in sensitivity (92% vs 80.4%, P < 0.001) and negative predictive value (94.4% vs 87.7%, P = 0.003), and it has also demonstrated marked improvement in the diagnostic capabilities of nonexpert endoscopists[49]. Additionally, several models were generated to identify the location and original layers of subepithelial lesions (SEL) with the use of endoscopic ultrasound (EUS). In a study by Hirai et al[50], images were utilized to train the AI system of pathologically confirmed the 5 major categories upper GI SEL: GI stromal tumor (GIST), leiomyoma, schwannoma, neuroendocrine tumor, and ectopic pancreas. Given that GIST is the most common type of SEL and are considered with potential for malignancy, the model utilized was able to differentiate GIST from non-GIST with specificity, sensitivity and accuracy of 98.8%, 67.6%, and 89.3%, respectively. With that study setting early precedent, a retrospective study utilized EUS images of esophageal SEL to develop a lesion detection algorithm called YOLOv8s-seg[51]. YOLO is real-time target detection algorithm that utilizes CNN for feature extraction and identification of species and location; however, this updated system provides better detection and segmentation while maintaining fast processing speed. The results of detection were then compared to both junior and senior endoscopists which is defined by cut-off value of at least 5 years of EUS experience and 500 EUS examinations, with kappa test to evaluate the consistency of the endoscopist’s diagnostic results. This YOLOv8s-seg model exhibited area under curve of 0.832 and significant diagnostic accuracy compared to junior endoscopist (76.5% vs 65.6%-66.7%, P = 0.045) due to the ability to detect lesions and identify the originating layer of the lesion[51] (Table 4).

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.

Gastric cancer

Gastric cancer has a high mortality risk with an estimated 769000 deaths annually where one out of every thirteen deaths is attributed to this condition, ranking fifth for incidence and fourth for mortality globally. It has a stronger predilection towards males with a two-fold increased risk that likely involves those residing in Eastern Asian and Eastern European countries[52]. The characterization of GC can be classified utilizing the Lauren Classification system and the World Health Organization (WHO) classification systems. Based on the Lauren Classification system of GC, it is separated into: Intestinal primary subtype, diffuse subtype, and mixed-type tumors exhibiting features of both intestinal and diffuse types. Intestinal type is characterized by gland formation with various degrees of differentiation, and the diffuse type is composed of poorly cohesive cells without gland formation that may contain signet ring cells[53]. The WHO classification system has 4 major histological patterns: Tubular, papillary, mucinous, and poorly cohesive/signet ring cell carcinoma[54]. There are a variety of causes where GC is attributed to including but not limited to Helicobacter pylori infection, Epstein Barr virus, familial causes such as germline mutations including hereditary diffuse GC due to E-cadherin gene mutation, dietary lifestyle with high red and processed meat intake, and smoking[55]. There are multiple different stages of GC which can be categorized utilizing the TNM system: “T” represents how far the tumor has grown in the stomach wall, “N” shows how cancer spread to nearby lymph nodes, and “M” demonstrates metastasis in which cancer has spread to other parts of the body[56].

AI incorporates ML and DL algorithms to assist with lesion detection during endoscopy and to analyze datasets such as clinical records, genomic and transcriptomic data, medical images, and digitized histopathology slides[57].

In a single-center, randomized controlled trial by Wu et al[46] in China, an AI system named ENDOANGEL-LD (lesion detection) was built upon the prior ENDOANGEL with a sensitivity of 91.5% and specificity of 88.1% for diagnosing neoplasm in a preliminary study in 5488 gastric images and real-time sensitivity of 100% and specificity of 54.84% in 100 upper GI endoscopy videos via the use of an augmented reality system. There is a relative reduction of 0.224 that has the ability to assist with lesion detection and reducing gastric neoplasm miss rates.

In a study by Gao et al[58] in 2024, incremental supervised contrastive learning model (iSCLM) included retrospective data of computed tomography scans and hematoxylin and eosin-stained biopsy images from 2387 patients in China. It was compared to incremental radiology-based model (iRM) and incremental pathology model (iPM) where the iSCLM has an area under the curve of 0.846 in the prospective cohort which is better than iRM (P = 0.007) and iPM (P < 0.001) to accurately predict the efficacy of neoadjuvant chemotherapy with the goal to improve outcomes.

A proposed model called GC risk assessment procedure with AI (GRAPE) utilizes three-dimensional noncontrast CT and DL to identify patients with GC with pixel-level segmentation outlining the stomach and potential tumors as well as integrating morphometric patterns and contextual radiological features. In a multicenter reader study, GRAPE surpassed radiologists in diagnostic accuracy with the area under the curve of 0.92 vs 0.76-0.85 in addition to higher proportion of early-stage, resected GC. This model reflected good performance on prediagnosis scans 6 months before GC diagnosis[59].

With the ability of AI for high-throughput sequencing and computations, liquid biopsy with real-time monitoring of tumor biomarkers has emerged as a methodology for early detection of GC such as the use of microRNAs (miRNAs) a class of non-coding RNA. The effective early-stage GC (ESGCmiRD) model identified 5 miRNAs (miR-320b, miR-222-3p, miR-181a-5p, miR-103a-3p, miR-107) with high diagnostic accuracy as it is overexpressed in the plasma of gastric carcinogenesis. This is accomplished by modulating the Hippo pathway that promotes tumor glycolysis in colorectal cancer, mediating p53 regulation of hypoxic signaling and angiogenesis, and affecting the PI3K/Akt pathway that targets PTEN via gene knockout to enhance proliferation, migration, and invasion of GC cells[60]. This is well studied utilizing miRNA expression from 10071 patients retrieved from Gene Expression Omnibus and The Cancer Genome Atlas databases.

In terms of utilizing AI and DL algorithm in the pathologic diagnosis of gastric tumors in endoscopic biopsies, Park et al[61] in 2021 developed an algorithm based on CNNs with 2434 whole-slide images that classified biopsy into three categories: Negative for dysplasia (NFD), tubular adenoma, or carcinoma. In this prospective study, the whole-slide image algorithm for GC detection had a 100% sensitivity, 97% specificity, 52% reduction in time for NFD cases (P < 0.001), and an area under the curve average of 0.9790. This model is cross-referenced and validated by comparing against the interpretation by six experienced pathologists with the limitation of the inability to detect nonepithelial tumors.

These aforementioned studies are examples of key AI models involving radiomics, pathomics, and multi-modal DL algorithms in its holistic approach to GC detection with specific applications. For instance, radiomics are best utilized to study metastasis detection, response prediction, and longitudinal monitoring while pathomics are optimal for time analysis for immunotherapy response, subtype classification, and toxicity prediction. The integration of both modalities allows for personalized therapy selection and prognostication[57]. Additionally, the use of AI in endoscopic imaging can assist with classification via the detection of early tumors, predicting invasion depth, and identifying precancerous lesions by magnifying images with flexible spectral imaging color enhancement and densely sampled scale-invariant feature transform descriptors utilizing multilayer CNN with multiple training and test images[62] (Table 5).

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.

Limitations

Despite the encouraging diagnostic performance reported across the studies reviewed, several recurrent methodological limitations warrant careful consideration before clinical translation. The vast majority of AI models were developed and validated using retrospective, single-center datasets — often enriched with near-equal proportions of diseased and non-diseased cases — introducing spectrum bias and inflating performance metrics beyond what would be expected in real-world screening populations where disease prevalence is substantially lower. Dataset heterogeneity in image acquisition protocols, annotation methodologies, endoscopic platforms, and outcome definitions further precludes meaningful cross-study comparison, while the geographic concentration of training data — predominantly from East Asian institutions — raises concerns about generalizability to Western populations with different disease epidemiology, endoscopic equipment, and clinical workflows. External validation, when performed, was frequently limited to geographically proximate centers with similar patient demographics, and reporting quality was often suboptimal, with many studies lacking confidence intervals, pre-registered protocols, calibration assessment, or risk-of-bias evaluation using validated tools. Finally, medico-legal questions regarding responsibility for AI-assisted diagnostic errors remain unresolved, emphasizing that current AI systems should function as decision-support tools rather than replacements for clinical judgment. Critically, very few studies employed prospective, randomized, controlled designs evaluating AI in real-time procedural environments. Until large-scale, multicenter randomized controlled trials with diverse populations, standardized benchmarking frameworks, and clinically meaningful endpoints are conducted, AI systems in upper GI endoscopy should be regarded as decision-support tools that augment expert clinical judgment rather than replace guideline-based endoscopic evaluation.

CONCLUSION

AI is rapidly emerging as a powerful adjunct in upper GI endoscopy. AI systems have demonstrated promising performance in identifying clinically relevant findings and improving the consistency of endoscopic interpretation across multiple applications, including lesion detection, risk stratification, and prognostication. Early studies in Barrett’s esophagus, gastric neoplasia, peptic ulcer disease, and GEV suggest that DL-based platforms can enhance diagnostic accuracy and assist in real-time decision-making during endoscopic procedures. Beyond diagnosis, AI-assisted findings may help guide subsequent clinical evaluation, histopathological confirmation, risk stratification, and surveillance planning. When integrated with established clinical guidelines and patient-specific factors, these insights may support more standardized and individualized management pathways.

Despite these advances, most AI models remain in early developmental stages and require prospective multicenter validation before widespread clinical adoption. Future research should prioritize prospective, multicenter studies evaluating AI-assisted detection of upper GI pathologies, as these applications have demonstrated encouraging diagnostic performance and are particularly suited to real-time endoscopic assessment. Studies should also incorporate diverse patient populations, standardized outcome measures, external validation, and clinically meaningful endpoints to determine whether improvements in diagnostic performance translate into better patient outcomes.

Ultimately, AI is likely to function as a decision-support tool that complements endoscopists’ expertise, with the potential to improve diagnostic accuracy, procedural efficiency, and patient outcomes in upper GI endoscopy. Its clinical utility will depend on the reliable integration of AI-generated findings into validated diagnostic and therapeutic pathways, while maintaining clinical judgment and multidisciplinary decision-making. Continued collaboration among clinicians, researchers, technology developers, and regulatory bodies will be essential to ensure that these technologies are validated rigorously and integrated safely and equitably into clinical practice.