Modern deep learning methods for predicting treatment unresponsiveness in idiopathic premature ventricular contraction patients

Background/Introduction Treatment unresponsiveness in idiopathic premature ventricular contractions (iPVC) represents a major clinical challenge with a substantial proportion of patients failing to achieve adequate PVC suppression despite antiarrhythmic therapy. Accurate risk stratification tools to predict treatment failure are lacking.Recent advances in deep learning (DL), particularly transformer-based architectures and novel neural networks such as the Kolmogorov-Arnold Network (KAN) may improve prediction modeling in structured clinical data. Purpose To develop and externally validate modern machine learning models for predicting treatment unresponsiveness in patients with iPVC and to compare their performance with conventional logistic regression. Methods This retrospective multicenter cohort included 1,299 patients.All patients had frequent iPVC (≥5% burden) without structural heart disease and received antiarrhythmic therapy for 2–6 months. Treatment unresponsiveness was defined as <80% reduction in PVC burden on follow-up Holter monitoring.Five models were developed: logistic regression, multi-layer perceptron (MLP), TabTransformer, TabNet, and KAN. Performance was evaluated using discrimination (AUC), calibration (intercept, slope, O:E ratio), Brier score and decision curve analysis. Logistic recalibration was performed and SHAP(SHapley Additive exPlanations)analysis was used for feature importance. Results Treatment unresponsiveness occurred in 31% of both cohorts. TabNet achieved the highest discrimination (AUC 0.982, 95% CI 0.970–0.989), followed by KAN (0.979, 0.957–0.990), TabTransformer (0.967, 0.942–0.981), MLP (0.965, 0.941–0.979) and logistic regression (0.925, 0.887–0.951).TabTransformer demonstrated the most balanced calibration, while KAN showed overestimation prior to recalibration. After recalibration, calibration improved across all models. Decision curve analysis showed higher net benefit for TabTransformer and KAN across clinically relevant probability thresholds.SHAP analysis consistently identified left ventricular ejection fraction, PVC burden, and PVC QRS duration as the most influential predictors. Higher LVEF was associated with lower risk of treatment unresponsiveness. Conclusion(s) Modern deep learning architectures outperform logistic regression in predicting treatment unresponsiveness in iPVC. Transformer-based models and KAN demonstrated superior discrimination and clinical utility, supporting their potential role in personalized decision-making and early identification of patients who may benefit from alternative strategies such as catheter ablation.