A lightweight SelfONN model for general ECG classification with pretraining

Abstract

Automatic ECG classification for arrhythmia detection is an import while challenging task in intelligent ECG analysis. Existing studies for ECG classification usually simplify the number of classes or only focus some of these categories, and a considerable part of it is for personalized (patient-specific) prediction or evaluated in intra-patient level (class-oriented), which brings challenges to the evaluation of models in terms of cross-individual generalization. SelfONN is a new biological neuron-inspired operator that is heterogeneous and has stronger nonlinear ability than convolution operation that as the basis of mainstream ECG classification algorithms. In this article, we propose a novel SelfONN-based model for general ECG classification. Besides, to leverage related large database to assist target-domain classification, we acquire pretrained models through pretraining, and in further explore different fine-tuning strategies. Furthermore, we design a target-guided, hierarchical classifier head to exploit the natural membership relationships between labels. Our proposed model (only about 45 K parameters with less than 1 M estimated total size) achieves macro-averaged AUC of 0.93 on the large PTB-XL database for a 71-classes multi-label classification task under inter-patient evaluation. Comparison with other mainstream methods shows that the proposed SelfONN model achieves competitive performance, and it requires fewer parameters and a smaller estimated total size, which is very convenient for further application and deployment. More broadly, this work shows the great potential of SelfONN beyond the commonly used traditional CNN architecture in ECG classification.