Subclinical Atrial Fibrillation Prediction in Patients with CIED by a Novel Deep Learning Framework

Abstract

Background: Subclinical atrial fibrillation (SCAF), a key risk factor for cryptogenic stroke, is difficult to predict with current tools. This study aimed to develop a novel deep learning framework, ResKAN-Attention, using only routine clinical data to predict SCAF in patients with cardiac implantable electronic device (CIED). Methods: In this retrospective study, the ResKAN-Attention model was developed using 27 routine parameters from 124 CIED patients without prior AF. This framework features a dual-path architecture combining a Kolmogorov–Arnold Network (KAN) with a traditional multilayer perceptron, fused via a cross-attention mechanism. The model’s performance was evaluated against common baselines using five-fold cross-validation, while its decision-making process was assessed through interpretability analysis. A clinically applicable risk scoring system was subsequently derived via knowledge distillation. Results: Over a 12-month follow-up period, SCAF occurred in 31.5% of patients (39/124). The ResKAN-Attention model significantly outperformed all baseline models, achieving a mean AUC of 0.837 in cross-validation and 0.788 in external validation. Interpretability analysis identified left atrial diameter (LAD), gender, lactate dehydrogenase, BMI, and hypertension as top predictors. The simplified risk score exhibited excellent predictive power (AUC 0.882), retaining 99.1% of the complex model’s performance on the fifth fold validation set. Conclusions: The ResKAN-Attention model demonstrated promising preliminary results for SCAF prediction with enhanced interpretability. The distilled risk score provided a potential method for early risk stratification in clinical settings, demonstrating that advanced artificial intelligence (AI) can effectively predict complex cardiovascular events using readily available data.