An interpretable shapelets-based method for myocardial infarction detection using dynamic learning and deep learning

Abstract Objective. Myocardial infarction (MI) is a prevalent cardiovascular disease that contributes to global mortality rates. Timely diagnosis and treatment of MI are crucial in reducing its fatality rate. Currently, electrocardiography (ECG) serves as the primary tool for clinical diagnosis. However, detecting MI accurately through ECG remains challenging due to the complex and subtle pathological ECG changes it causes. To enhance the accuracy of ECG in detecting MI, a more thorough exploration of ECG signals is necessary to extract significant features. Approach. In this paper, we propose an interpretable shapelet-based approach for MI detection using dynamic learning and deep learning. Firstly, the intrinsic dynamics of ECG signals are learned through dynamic learning. Then, a deep neural network is utilized to extract and select shapelets from ECG dynamics, which can capture locally specific ECG changes, and serve as discriminative features for identifying MI patients. Finally, the ensemble model for MI detection is built by integrating shapelets of multi-dimensional ECG dynamic signals. Main results. The performance of the proposed method is evaluated on the public PTB dataset with accuracy, sensitivity, and specificity of 94.11%, 94.97%, and 90.98%. Significance. The shapelets obtained in this study exhibit significant morphological differences between MI and healthy subjects.