QRS-centric beat-wise atrial fibrillation detection in ECG signals using deep neural networks

We propose a deep learning approach for beat-wise atrial fibrillation (AF) detection in electrocardiogram (ECG) signals. AF, a major cardiac arrhythmia affecting millions globally, requires early detection for optimal treatment outcomes. Current rhythm-based methods lack beat-level precision and timeliness. Our QRS-centric approach improves traditional classification by providing precise beat-wise analysis, addressing the critical gap between rhythm-level detection and immediate arrhythmia identification. Our network architecture integrates convolutional layers for feature extraction with bidirectional LSTM layers that target the QRS complex region through adaptive beat segmentation. This architecture enables detailed temporal dependency analysis while preserving morphological feature sensitivity. Extensive evaluation on the MIT-BIH Arrhythmia Database demonstrates exceptional performance with 98.606% precision and 99.648% sensitivity for AF detection. The model's robust generalizability is confirmed through independent validation on the MIT-BIH Atrial Fibrillation Database, achieving a 98.653% F-score. Further validation using contemporary datasets (MIMIC-III and Simband) confirms the model's effectiveness in identifying multiple beat types including normal sinus rhythm, AF, premature ventricular contraction, and premature atrial contraction. The QRS-centric approach enables timely arrhythmia detection at the individual beat level, demonstrating a significant advancement in automated ECG analysis. The high accuracy and beat-level precision of this approach suggest potential for integration with wearable devices and real-time monitoring systems, enabling earlier clinical intervention and improved arrhythmia diagnosis in practice.