Lightweight Optimization of Deep Learning Models for Accurate Arrhythmia Detection in Clinical 12-Lead ECG Data

Although significant progress has been made in diagnosing arrhythmias, current deep neural network (DNN)-based methods still suffer from misdiagnosis and redundant parameters. One reason for this is the lack of high-quality clinical 12-lead electrocardiogram (ECG) datasets and the absence of a systematic set of general task-oriented network lightweight methods. The objective of this article is to propose a lightweight optimization method of deep learning models for accurate arrhythmia detection in clinical 12-lead ECG data. We show that the proposed optimized module replacement method can provide a strategy for arrhythmia detection in the lightweight network design for clinical 12-lead classification. The key of this method is to design compact convolution to replace standard convolutions based on module functional equivalence and identify the accuracy fractured segment (AFS). To achieve this, we randomly choose a network base (the initial model) that achieves satisfactory classification performance on a given classification task, replace larger-scale parameter convolutions with smaller-scale parameter convolutions, observe the degenerated process of accuracy until an AFS is found, and conduct tests in the candidate models. We validate the proposed method against the three-year 12-lead ECGs of patients who have undergone ECG detection at Shanghai First People’s Hospital and achieve an overall accuracy of 95.532%, a precision score of 0.927, a recall score of 0.910, and a specificity score of 0.990, which are better than other methods. The comparative analysis shows that the proposed method can not only effectively reduce network parameters but also boost model performance for specific classification tasks.