Enhancing Atrial Fibrillation Prediction Using Multimodal Neural Networks and XGBoost Integration

Atrial fibrillation (AF), a prevalent cardiac arrhythmia, poses significant health risks, emphasizing the critical need for accurate and timely detection to develop robust methodologies and ensure effective intervention. This study propels the field of AF detection through detailed electrocardiogram (ECG) analysis, employing sophisticated deep learning models and multimodal frameworks aimed at leveraging the complementary strengths of both deep learning and additional data sources, including heart rate variability (HRV) features and demographic data. The potential enhancement from the integration of high-level features from a deep neural network with XGBoost is also assessed. Our evaluation involves 35,634 recordings from three diverse public 12-lead ECG databases (CPSC-Extra, PTB-XL, and Georgia) with physician-interpreted AF labels, revealing notable performance disparities among various CNN architectures. AlexNet and VGG-16, less complex models, exhibited high performance both individually and in multimodal settings. Specifically, AlexNet achieved 96.80% AUC-ROC individually and 97.23% in multimodal modeling, while VGG-16 reached 97.13% in multimodal approach. The inclusion of HRV parameters positively impacted performance, while demographic data showed limited influence. Integrating high-level features with XGBoost increases overall accuracy but reduces AUC-ROC; for AlexNet, an AUC-ROC of 90.92% was observed. This study offers valuable insights into optimal model configurations, emphasizing the pivotal impact of a multimodal framework.