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

Area of Science:

• Artificial Intelligence in Cardiovascular Medicine
• Machine Learning for Predictive Diagnostics
• Biomedical Data Science

Background:

• Subclinical atrial fibrillation (SCAF) is a significant risk factor for cryptogenic stroke.
• Current prediction tools for SCAF in patients with cardiac implantable electronic devices (CIEDs) are limited.
• Routine clinical data holds potential for predicting SCAF.

Purpose of the Study:

• To develop a novel deep learning framework, ResKAN-Attention, for SCAF prediction.
• To utilize only routine clinical data for SCAF prediction in CIED patients.
• To create an interpretable and clinically applicable risk scoring system.

Main Methods:

• Development of the ResKAN-Attention model using 27 routine parameters from 124 CIED patients.
• A dual-path architecture combining Kolmogorov-Arnold Network (KAN) with a multilayer perceptron, fused via cross-attention.
• Performance evaluation using five-fold cross-validation and comparison with baseline models; interpretability analysis and knowledge distillation for risk score derivation.

Main Results:

• SCAF incidence was 31.5% (39/124) over 12-month follow-up.
• ResKAN-Attention significantly outperformed baseline models (AUC 0.837 cross-validation, 0.788 external validation).
• Key predictors included left atrial diameter, gender, lactate dehydrogenase, BMI, and hypertension; a simplified risk score achieved high predictive power (AUC 0.882).

Conclusions:

• The ResKAN-Attention model shows promise for SCAF prediction with enhanced interpretability.
• The derived risk score offers a potential tool for early risk stratification in clinical practice.
• Advanced AI can effectively predict complex cardiovascular events using readily available clinical data.