Patient-Adaptive Beat-Wise Temporal Transformer for Atrial Fibrillation Classification in Continuous Long-Term Cardiac Monitoring

Atrial fibrillation (AF) is a prevalent cardiac arrhythmia that requires accurate diagnosis and management, especially in long-term cardiac monitoring (LTCM) scenarios. Although ECG signal morphology can vary between patients and over time, traditional deep learning models often focus on generalized classification, potentially overlooking patient-specific differences. In this study, we developed a patient-adaptive beat-wise temporal transformer model aimed at enhancing AF classification performance using data from LTCM devices. The model introduces a symbolic token representing the ideal ECG morphology of a patient, enabling the transformer to effectively reference patient-specific information and capture temporal variations in ECG morphology. To evaluate the model’s classification performance, we trained it on public databases and tested it on patch datasets. The model achieved an accuracy of 0.987, precision of 0.965, sensitivity of 0.979, specificity of 0.99, and an F1 score of 0.982 on the Patch A dataset. On the Patch B dataset, it attained an accuracy of 0.953, precision of 0.91, sensitivity of 0.971, specificity of 0.942, and an F1 score of 0.951. We anticipate that the proposed approach will be particularly beneficial for monitoring AF in real-world clinical settings, especially in LTCM scenarios.