Objective To investigate morphological differences between obstructive and non-obstructive coronary artery disease (CAD) patients using computer-aided image analysis, and identify color and texture features for traditional Chinese medicine (TCM) syndrome differentiation.

Methods This prospective study enrolled patients undergoing coronary computed tomography angiography (CTA) at the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine between May 1, 2024 and August 7, 2025. Based on CTA results, patients were categorized into obstructive CAD and non-obstructive CAD groups. Standardized tongue images were acquired using a dedicated mobile application (Traditional Chinese Medicine Tongue Image-Assisted Diagnosis System) and analyzed for the overall tongue surface and three macroscopic features (tooth marks, fissures, and red dots) from which high-dimensional color and texture parameters were extracted. Multi-scale texture features were derived using spatial-domain Laplacian pyramid and frequency-domain wavelet transform methods. Dimensionality reduction and feature selection were performed using principal component analysis (PCA) and random forest with 5-fold cross-validation. Feature stability was assessed using Hodges-Lehmann estimator and Cliff’s δ. A multi-view XGBoost model was developed to differentiate the two groups and evaluated on a temporally independent validation set using accuracy and the area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) analysis was applied to interpret model decisions.

Results This study analyzed 373 CAD patients, including 167 with obstructive CAD and 206 with non-obstructive CAD according to CTA results. The whole cohort was divided into training set (n = 316, obstructive : non-obstructive = 142 : 174 ) and validation set (n = 57, obstructive : non-obstructive = 25 : 32), with balanced baseline characteristics (P > 0.05). Macroscopic tongue analysis revealed that patients with obstructive CAD had fewer tooth marks odds ratio (OR) = 0.43, P < 0.05 and red dots (OR = 0.46, P < 0.05). High-dimensional color analysis identified pronounced intergroup differences, most notably a reduction in hue values in the hue-saturation-intensity (HSI) color space among obstructive CAD patients (Cliff’s δ = – 0.31, P = 2.72 × 10^–6; Hodges-Lehmann estimator: – 0.31). PCA results suggested that tongue surface features explained the highest proportion of variance (48.2%). Random forest screening identified 77 stable features across all tongue regions, with wavelet-transformed texture features demonstrating the highest importance. The multi-view XGBoost fusion model achieved an accuracy of 75% and an AUC of 0.779 in the independent validation set. SHAP analysis identified the wavelet-based feature—left-handed lower-level gray-level size zone matrix zone variance (LHL_glszm_ZoneVariance) as the top predictor, accounting for 40.6% of the model's decision variance, and indicated that 85.3% of the predictive power came from wavelet-based texture features.

Conclusion This study has provided objective evidence for the TCM concept that “the tongue reflects the heart” by identifying distinct morphological and colorimetric tongue patterns in patients with obstructive CAD through artificial intelligence (AI)-driven image analysis, and the promising performance of the computational model suggests its potential as a non-invasive adjunctive tool for CAD assessment.