An interpretable machine-learning model for Traditional Chinese Medicine syndrome differentiation in ischemic stroke using integrated disease–syndrome features

Objective To develop and internally validate an interpretable machine-learning model for traditional Chinese medicine (TCM) syndrome differentiation in ischemic stroke by integrating modern clinical characteristics with TCM-related features. Methods A total of 616 patients with ischemic stroke from a tertiary TCM hospital in Sichuan Province, China, were included. Fifty-one candidate predictors covered demographics, stroke severity, laboratory indices, infarction characteristics, medical history, TCM symptoms, tongue features, and pulse characteristics. Five TCM syndrome classes were considered. Patients were stratified into a training set (n = 495) and a held-out test set (n = 121). Grouped multinomial least absolute shrinkage and selection operator regression was used for feature selection. Random forest (RF), support vector machine with a radial basis function kernel (SVM-RBF), and extreme gradient boosting (XGBoost) were trained using identical stratified 10-fold cross-validation partitions. Performance was assessed using accuracy, Cohen’s κ, Macro-F1, macro balanced accuracy, Macro AUC, and Hand-Till multiclass AUC, with 95% confidence intervals estimated by 2,000 bootstrap resamples. Model calibration and TreeSHAP-based interpretability were further evaluated for the final model. Results At λ1se = 0.0281, 33 encoded features were retained. In the held-out test set, RF achieved the highest accuracy (0.752; 95% CI, 0.678–0.826) and Macro-F1 (0.744; 0.650–0.819), whereas XGBoost achieved the highest Macro AUC (0.945; 0.916–0.970) and Hand-Till multiclass AUC (0.944; 0.912–0.970). Overall classification correctness did not differ significantly among the three models (Q = 4.065, P = 0.131), and paired RF–XGBoost bootstrap differences showed no clear overall performance advantage. XGBoost was therefore selected for further analysis. Its multiclass Brier score was 0.358 (0.264–0.463), with a Brier skill score of 0.540. SHAP identified LDL-C, age, wiry pulse, BMI, waist and knee soreness, red tongue, enlarged tongue, homocysteine, triglycerides, and hypertension as leading global contributors, with distinct class-specific attribution patterns across syndromes. Conclusion XGBoost showed promising discrimination in the single-center internal test set; however, classification performance and calibration varied across syndrome classes. These findings represent a preliminary internal evaluation and do not establish clinical readiness or support the use of the model for clinical decision-making or prescription guidance. External validation and prospective evaluation are required before clinical implementation.