Artificial intelligence-based tongue image analysis for non-invasive screening of gastric cancer, precancerous lesions, and related gastrointestinal disorders: a systematic review

BACKGROUND: Gastric cancer remains a major cause of cancer-related mortality, largely because many cases are diagnosed at an advanced stage. Upper gastrointestinal endoscopy with histopathological confirmation is the diagnostic gold standard, but its invasiveness, cost, and limited accessibility restrict population-level use. Artificial intelligence (AI)-based tongue-image analysis may provide an objective, low-cost, and non-invasive adjunct for pre-endoscopic risk stratification. METHODS: This systematic review was conducted according to PRISMA 2020 principles and was informed by diagnostic test accuracy and prediction-model reporting frameworks. PubMed/MEDLINE, Scopus, Web of Science, and IEEE Xplore were searched from inception to 10 June 2026. Eligible studies were original investigations applying machine-learning or deep-learning methods to digital tongue images for gastric cancer, precancerous gastric lesions, or related gastrointestinal disorders and reporting at least one quantitative performance metric. Data were extracted on study design, sample size, target condition, reference standard, image-acquisition protocol, model architecture, validation strategy, diagnostic performance, calibration, interpretability, and risk of bias. A narrative synthesis was performed because of substantial heterogeneity. RESULTS: The search identified 139 records, of which 78 remained after duplicate removal. After title, abstract, and full-text screening, 9 studies published between 2018 and 2025 were included. The evidence covered gastric cancer, precancerous gastric lesions or high-risk gastric screening, and related gastrointestinal disorders. Models included logistic regression and XGBoost classifiers based on handcrafted features, convolutional neural networks, transformer-based architectures, segmentation-assisted pipelines, and multimodal models combining tongue images with clinical, questionnaire, biomarker, or microbiome data. Internally validated AUCs ranged from 0.61 to 0.99 across target conditions. External validation was reported in only three studies and generally showed performance degradation, particularly for precancerous gastric lesions. Calibration assessment was uncommon, and most studies did not report calibration curves, Brier scores, expected calibration error, or recalibration procedures. CONCLUSION: AI-based tongue-image analysis is a promising but early-stage approach for non-invasive gastric disease screening. Current evidence supports its potential role as an adjunctive pre-endoscopic triage tool rather than a replacement for gastroscopy. Prospective external validation, standardized image acquisition, transparent model reporting, calibration assessment, and evaluation in diverse populations are required before clinical deployment. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420261444857, PROSPERO CRD420261444857.