Artificial intelligence and digital transformation of gastroenterology and hepatology: A critical review of clinical applications and future challenges

Table 1 Summary of artificial intelligence applications for non-invasive assessment of fibrosis and cirrhosis

Input data type	AI tool	Conventional metric/tool outperformed	Key AI contribution	Evidence level	Primary limitation
Clinical/Laboratory	RF, LightGBM, ANN	FIB-4 score and transient elastography	More reliable prediction of fibrosis stage; development of novel indices (e.g., FIB-6, validated in multicenter cohorts)	Moderate-high (multicenter validation available for FIB-6; large cohorts in MASLD studies) near clinical use	Some models remain retrospective; limited external validation for several algorithms; interpretability constraints
Imaging (CT, MRI, US, elastography)	DL (CNNs, ResNet50), Radiomics	Expert radiologists; elastography alone	Enhanced cirrhosis detection; automatic segmentation; identification of inflammation/fibrosis; distinction of etiologies; radiomics improves staging precision	Moderate (several studies with external validation, but heterogeneous datasets), promising	Data heterogeneity; many single-center cohorts; limited standardized imaging protocols; 'black box' interpretability
ECG	DL (AI- cirrhosis-ECG score)	Standard clinical evaluation	Low-cost cirrhosis screening with high AUC (0.908); potential for routine, scalable screening	Low-moderate (retrospective, single-center), emerging	Limited sample size; lack of external validation; implementation barriers despite low test cost

RF: Random forest; ANN: Artificial neural networks; CT: Computerized tomography; MRI: Magnetic resonance imaging; DL: Deep learning; CNN: Convolutional neural networks; AUC: Area under the curve; AI: Artificial intelligence; LightGBM: Light Gradient Boosting Machine; ECG: Electrocardiography; MASLD: Metabolic dysfunction-associated steatotic liver disease; FIB-4: Fibrosis-4; FIB-6: Fibrosis-6.

Table 2 Application of artificial intelligence in hepatocellular carcinoma diagnosis

Management stage	Data source/technology	AI application	Key benefit	Evidence level	Primary limitation
Risk stratification	Clinical, viral (HCV), cirrhosis data	ML models (outperform traditional regression)	Superior prediction of HCC risk, even post-HCV eradication or in MASLD	Research	Retrospective design; single center; limited generalizability; data heterogeneity; 'black box' nature
Imaging diagnosis	Radiomics (CT, MRI)/DL	Classification of benign vs malignant lesions	Identifies indeterminate nodules and predicts MVI	Research	Validation needed in diverse populations; 'black box' interpretability; generalizability across centers
Pathology	H&E histological slides	DL/CNNs	Classification of tumor subtypes and prediction of genetic mutations with near-perfect reliability	Research	Dependence on quality of scanned slides; potential for algorithmic bias; 'black box' interpretability
Prognosis/treatment	Imaging/multi-omics/clinical data	Estimation of survival and therapeutic response	Prediction of recurrence and response to locoregional therapies (e.g., TACE)	Research	Need for broader multicenter validation; data heterogeneity; 'black box' nature; integration into clinical workflow

HCV: Hepatitis C virus; ML: Machine learning; CT: Computerized tomography; MRI: Magnetic resonance imaging; DL: Deep learning; MVI: Microvascular invasion; H&E: Hematoxylin and eosin; CNN: Convolutional neural networks; TACE: Transarterial chemoembolization; AI: Artificial intelligence; HCC: Hepatocellular carcinoma.

Table 3 Relevant applications of artificial intelligence in inflammatory bowel disease

Application area	Data source and type	AI application/technique	Key clinical benefit	Evidence level	Primary limitation
Non-invasive diagnosis	Fecal multi-omics	ML models	Accurate differentiation of healthy vs UC vs CD	Research/early clinical	Data heterogeneity, limited external validation, small/retrospective datasets, lack of generalizability
Differential diagnosis	Endoscopic imaging and clinical records (NLP)	TextCNN/image analysis	Distinguishes CD from intestinal tuberculosis and UC from CD	Research	Symptoms overlap and endoscopic similarities; limited data quality; need for external validation
Disease assessment	Radiomics and endoscopic video	DL models	Quantifies inflammation, detects strictures, and automates endoscopic activity scoring improving standardization	Research	Limited data quality and standardization; lack of external validation; 'black box' nature
Histological prediction	Histopathological slides	CNNs/DL (automating RHI, NHI, PHRI)	Objective scoring and superior prediction of future flares and post-surgical recurrence	Research	High inter- and intra-observer variability in expert labeling; data quality; 'black box' interpretability
Therapeutic response	Clinical, laboratory, multi-omics, endoscopy data	ML predictive models	Predicts response to biologic treatment, enabling timely therapy adjustment	Research/early clinical	Disease heterogeneity; variable treatment responses; need for robust external validation; 'black box' interpretability
Risk stratification	Peripheral blood transcriptomics, histology	ML/DL models (low vs high-risk groups)	Predicts disease progression, need for treatment escalation, and post-surgical recurrence/complications (strictures/fistulas). Recurrence/complications (strictures/fistulas)	Research	Need for larger, diverse datasets; limited external validation; potential algorithmic bias
Drug development	Transcriptomic data (intestinal tissue)/Boolean networks	Identification of novel therapeutic targets	Accelerates the discovery of first-in-class therapies with novel mechanisms of action	Research	Complexity of biological systems; need for robust preclinical validation; 'black box' nature; ethical considerations
Clinical trials	Patient electronic medical records	NLP/DL algorithms	Accelerates patient recruitment and potential use of "Digital Twins"	Research/future prospect	Data privacy and security; regulatory frameworks; ethical concerns regarding "digital twins" and placebo groups

UC: Ulcerative colitis; CD: Crohn’s disease; NLP: Natural language processing; DL: Deep learning; CNN: Convolutional neural networks; RHI: Robarts Histopathology Index; NHI: Nancy Histological Index; PHRI: PICaSSO Histologic Remission Index; AI: Artificial intelligence; ML: Machine learning.

Table 4 Role of computer-aided detection technology in endoscopy

Technology	Function	AI action	Clinical output	Impact	Evidence level	Primary limitation
CADe (detection)	Real-time lesion localization	Processes endoscopic video stream and automatically highlights suspicious regions with visual overlays	Increased detection rate	Prevents missed lesions (e.g., sessile serrated lesions, early gastric cancer), particularly for less experienced endoscopists	Clinical trial/commercial	High rate of false positives leading to endoscopist fatigue; potential for increased procedure time; lack of generalizability across diverse populations
CADx (diagnosis)	Real-Time lesion characterization	Analyzes morphological and vascular patterns of a detected lesion	Classification of lesions (adenomatous vs non-adenomatous) or activity grading in IBD	Facilitates "Resect and Discard" or "Diagnose and Leave" strategies, reducing biopsy costs and time	Clinical trial/commercial	Variability in accuracy based on polyp location (e.g., proximal vs distal colon); 'black box' nature limiting clinician trust; need for external validation
Integrated system	Informed therapeutic decision	Simultaneous use of CADe and CADx in a single, non-interruptive workflow	Precise treatment plan	Optimizes workflow efficiency and standardizes quality of care across different operators	Clinical trial/commercial	Seamless integration into diverse healthcare environments; high development/maintenance costs; 'deskilling' risk for endoscopists; unclear regulatory/ethical frameworks for AI-assisted decisions

IBD: Inflammatory bowel disease; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; AI: Artificial intelligence.

Table 5 Evidence levels across major artificial intelligence applications in gastroenterology and hepatology

Domain/application	Evidence level	Key determinants
Endoscopy (CADe/CADx)	High	Multiple FDA/CE-approved tools; prospective multicenter trials; real-time clinical use
Radiology (CT/MRI)	Moderate-high	External validation common; some multicenter cohorts; radiomics + DL pipelines
Non-invasive liver tests/fibrosis	Moderate	Mix of large cohorts + retrospective datasets; limited external validation except for FIB-6
HCC detection and surveillance	Low-moderate	Early-stage models; heterogeneous metrics; mostly retrospective; few external validations
IBD (imaging, histology)	Moderate	Prospective validation for endoscopy/histology models; omics models still experimental
IBD multi-omics/transcriptomics	Low	Experimental; small cohorts; no external validation
Capsule endoscopy	Moderate	Strong DL performance; mostly single-center retrospective datasets
Motility testing/manometry	Low-moderate	Emerging field; small datasets; experimental DL approaches
Predictive models for complications	Low-moderate	Mostly retrospective; internal validation only
Implementation/regulation	-	Not applicable (conceptual section)

IBD: Inflammatory bowel disease; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CT: Computerized tomography; MRI: Magnetic resonance imaging; DL: Deep learning; CE: Capsule endoscopy; HCC: Hepatocellular carcinoma.