Atrial Fibrillation Identification Using CNNs Based on Genomic Data

Abstract

Atrial fibrillation (AF) is the most common cardiac arrhythmia and a major cardiovascular disease epidemic of the 21st century. Early diagnosis and intervention are crucial as AF often progresses without symptoms. This study aims to identify AF using genome-wide association studies and convolutional neural networks (CNN). Genomic data from 6,358 individuals were used to develop a CNN model, with L2 regularization applied to prevent overfitting. The L2-regularized CNN significantly outperformed the regular CNN across various p-value thresholds. For instance, at p < 0.0001, the L2-regularized CNN achieved an accuracy of 0.731 ± 0.071 compared to 0.703 ± 0.055 for the regular CNN. At p < 0.001, the L2-regularized CNN showed an accuracy of 0.630 ± 0.089, while the regular CNN had 0.577 ± 0.095. This demonstrates a notable improvement in model performance with L2 regularization. Although the regular CNN showed higher accuracy in some scenarios, such as achieving 0.984 ± 0.015 at p < 0.01 compared to 0.970 ± 0.020 for the L2-regularized CNN, the performance difference between the models decreased as the p-value threshold became more stringent. Overall, L2 regularization not only improved the model’s performance and stability but also reduced the performance gap between the models under stricter p-value conditions. These findings highlight that L2-regularized CNNs can significantly enhance performance in genomic studies, offering a more effective alternative to traditional polygenic risk score methods for AF identification study.