QRS-Centric Deep Learning for Precision Beat-wise Atrial Fibrillation Detection in ECG Analysis

We propose a multi-stage deep learning approach for heartbeat-wise detection of atrial fibrillation (AF) from electrocardiogram (ECG) signals. Our method combines sample-level classification with beat-level aggregation, culminating in a QRS complex-focused analysis that aligns closely with clinical ECG interpretation practices. The model utilizes a deep neural network architecture featuring 1-D convolutional layers for feature extraction and bidirectional long short-term memory (BiLSTM) layers for capturing temporal dependencies. We trained and validated the model using carefully selected records from the MIT-BIH Arrhythmia Database. Our approach demonstrates exceptional performance in beat-wise AF detection, particularly in the final QRS-focused stage. On the independent test set, we achieved an overall accuracy of 99.80% for AF detection. The model showed near-perfect precision, sensitivity, and specificity for both AF and normal beat classification, with AF detection showing 1.0 precision, 0.9981 sensitivity, and 1.0 specificity. The key innovation of our work lies in the progressive refinement of the classification process, from sample (time-stamp) level to beat-wise, and ultimately focusing on the QRS complex region. This stepwise approach significantly enhances detection accuracy. By concentrating on the QRS complex, our method captures the most relevant information for distinguishing between normal and AF beats, improving both accuracy and physiological relevance of the classification process. This work has significant implications for improving the accuracy of AF diagnosis in clinical settings, potentially leading to more timely interventions and improved patient outcomes.