Patient-Specific Explainable Arrhythmia Detection Using Attention-Based Deep Networks

The aim of this study is to provide a patient-centered and explainable deep learning framework for arrhythmia detection from multi-lead electrocardiogram (ECG) signals, a framework that, in addition to improving classification performance, can provide meaningful interpretive evidence to support clinical decision-making. In this study, the PTB-XL dataset was used as the main data source, and a complete process, including patient-wise data partitioning, signal preprocessing, target label construction, and multi-label classification, was designed. The proposed model was developed based on learning patient-specific representations and used an attention-based deep network to extract temporal dependencies and lead-wise importance in the ECG signal. The key idea is to combine information from the current ECG recording with patient-specific representations to provide a more patient-aware prediction. Additionally, an ablation study and explainability analysis were performed to evaluate the role of architectural components. In the hold-out test on the evaluation dataset, the proposed framework demonstrated an area under the precision–recall curve (AUPRC) of 0.712. According to the comparative results, an increase in themacro-F1 score was observed from 0.681 to 0.703, that is, an improvement of 0.022 points. For the explainability assessment, the Peak Overlap value for the atrial fibrillation (AFIB) class was bout 89.6%. The primary contribution of the current study lies in the development of an integrated framework that combines patient-specific representationlearning, temporal attention, lead-wise attention, and multimodal fusion. However, the obtained results should be carefully considered since the proposed model was tested retrospectively on one publicly available dataset without any external prospective or clinician-based clinical validation.