Morphology-Aware ECG Diagnostic Framework With Cross-Task Attention Transfer for Improved Myocardial Infarction Diagnosis

In recent years, deep learning has been widely used in automated interpretation of electrocardiogram (ECG) signals, and several high-performing models have been developed. However, a majority of the existing deep learning models adopt a data-driven approach agnostic of any clinical priors. In this work, we present a novel morphology-aware diagnostic model for the detection of myocardial infarction (MI) from ECG signals. Clinically, MI is associated with specific morphological alterations in the ECG waveform. Integration of such clinical context into the learning process can aid in diagnosis. To facilitate this, we first design an attention-based ECG delineation network, focusing on learning morphology-specific attention. Then, we introduce a cross-task attention transfer (CTAT) framework to transfer the morphology-specific information of the delineation task to the MI detection network, thereby incorporating morphological context into the MI detection task. We evaluate our method on three public 12-lead ECG datasets. Experimental evaluation shows that our proposed framework achieves promising performance for MI detection from 12-lead ECG signals. Results also demonstrate superior performance compared with other state-of-the-art methods. Moreover, analysis of the post hoc interpretation in the form of Grad-CAM attributions verifies the model’s performance and provides clinically meaningful inferences to its predictions.