Abstract
BACKGROUND: Cardiovascular events (CVEs) are the leading cause of mortality in hemodialysis patients. Current prediction models rely on clinical and biochemical data, but non-invasive alternatives are needed. Inspired by the Traditional Chinese Medicine (TCM) principle that "the heart opens into the tongue," this study investigated whether quantitative features from tongue images could enhance CVE prediction. OBJECTIVE: To develop and validate a machine learning framework that integrates tongue image features with conventional clinical variables to predict CVEs in hemodialysis patients. METHODS: In this prospective, multicenter study, 506 maintenance hemodialysis patients were recruited. We extracted 1,354 hand-crafted radiomic features and 8 deep-learning features from standardized tongue images. These were combined with 90 clinical variables. Using a dataset split into training (n=243), validation (n=105), and an independent external test set (n=158), we developed and compared four models (LR, LightGBM, AdaBoost, MLP) under three feature configurations: clinical-only, tongue-only, and a fused model. RESULTS: The model using only tongue image features (AdaBoost) significantly outperformed the clinical-only model, achieving an AUC of 0.786 vs. 0.682 on the external test set. The fused model provided a marginal improvement (AUC=0.787). SHAP analysis indicated that both tongue texture features and clinical biomarkers like PT% were key predictors. Decision curve analysis confirmed the clinical utility of the tongue-based and fused models across a range of risk thresholds. CONCLUSION: Tongue image features are potent, non-invasive predictors of CVEs in hemodialysis patients, offering performance superior to conventional clinical variables. This AI-driven approach validates the TCM theory and presents a promising supplementary tool for enhancing risk stratification in nephrology care.