Abstract
OBJECTIVES: The high rate of rehospitalization following catheter ablation in atrial fibrillation (AF) patients remains a significant clinical challenge. This study aimed to develop a novel prediction model based on Kolmogorov-Arnold Networks (KANs) for postoperative rehospitalization and explore its clinical potential. Additionally, interpretability methods were employed to identify key risk factors. METHODS: Real-world clinical data from 430 AF patients who underwent catheter ablation were collected. Core predictors were selected through feature engineering. A KANs-based prediction model was constructed and compared with seven traditional machine learning models, including Support Vector Machines and Random Forests. Model performance was systematically evaluated using metrics such as accuracy and recall. The SHapley Additive exPlanations framework was applied to interpret feature contributions and conduct individual case analyses. RESULTS: The KANs model demonstrated superior predictive performance, achieving an area under the curve of 0.85, representing a 12% improvement over the suboptimal model. Key predictors included Low-Density Lipoprotein Cholesterol and Total Cholesterol. Individual case analyses revealed that the model effectively identified high-risk patients through biochemical indicator patterns, enhancing its interpretability. CONCLUSIONS: This study is the first to validate KANs in predicting postablation rehospitalization, enabling precise predictions and identifying critical biomarkers, thereby laying the foundation for improved postoperative management.