MT-MV-KDF: A novel Multi-Task Multi-View Knowledge Distillation Framework for myocardial infarction detection and localization

Abstract

Myocardial infarction (MI) is a prevalent and life-threatening cardiovascular ailment, diagnosed primarily through the twelve-lead electrocardiogram (ECG). These leads provide unique insights into electrical activity across diverse heart regions, crucial for MI differentiation and assessment. However, prior studies predominantly focus on single-perspective 12-lead ECG analysis, overlooking inter-lead disparities. Moreover, prevailing methods often lack versatility, being tailored to specific tasks. Therefore, this paper presents the Multi-Task Multi-View Knowledge Distillation Framework (MT-MV-KDF) for MI detection and localization. The framework combines multi-view learning, multi-task learning, and knowledge distillation to enhance model performance and efficiency. Specifically, The introduction of multi-view learning aims to learn ECG feature representations from different views. Additionally, multi-task learning enables the model to learn the similarities and differences of multiple related tasks simultaneously. Finally, knowledge distillation is used to transfer latent knowledge from the complex teacher model to the simpler student model to improve efficiency. The MT-MV-KDF consists of MT-MVT-net and MT-SVS-net, which use a multi-layer CNN architecture to extract ECG different scale features. An attention module is used to capture inter-channel information and spatial cues. Through knowledge distillation, the MT-DSVS-net (0.40M) inherits insights from MT-MVT-net while achieving comparable performance with fewer parameters. Evaluation on the PTB-XL database demonstrates MT-DSVS-net has an Acc, AUC of 93.78% and 92.23% for MI detection and 83.45% and 78.26% for MI localization, respectively. Additionally, Robustness validation on CPSC2018 and Chapman datasets further confirms the efficacy of MT-MV-KDF. This study highlights the MT-MV-KDF effectiveness in MI detection and localization, particularly in improving inter-patient MI localization.