Introduction: Coronary artery lesions (CALs) represent the most severe complication of Kawasaki disease (KD). Existing prognostic tools mainly assess overall coronary artery involvement and lack the ability to identify different grades of CAL at an early stage, especially in terms of bedside prediction tools based on artificial intelligence. (AI). Objectives: This study aimed to establish and externally validate an AI-driven bedside model capable of predicting varying degrees of CALs in children with KD. Methods: A stacking model integrating 11 base learners was developed using 2,696 KD cases from Children's Hospital of Soochow University (Sept 2020–Sept 2024), split 4:1 for training/internal validation. The model was evaluated via accuracy , precision, recall, F1-score, and AUC, and validated externally across three independent hospitals. SHAP values were used to interpret feature contributions. Results: Thirty-eight predictive variables were identified and analyzed. Among the tested algorithms, ExtraTrees, CatBoost, and AdaBoost achieved the highest AUCs and MCCs for normal, mild, moderate, and severe dilation, respectively , indicating consistent discriminative capability across CAL categories with AUCs of 0.819, 0.788, 0.798, and 0.704. SHAP analysis quantified global feature importance and enabled individualized interpretation of model outputs. Based on these variables, an intuitive bedside tool was constructed to estimate the risk of CAL with varying severity in children with KD (https://wszx133predict.streamlit.app/). Conclusions: This study establishes the first interpretable ML framework for differentiating KD CAL severity using routine data. Featuring strong multi-center generalizability and a webbased platform for real-time predictions, it enhances early risk recognition to potentially reduce coronary complications.