Optimizing Neural Network Scale for ECG Classification

TL;DR

This work investigates how to optimally scale CNNs, specifically ResNet architectures, for ECG classification by evaluating key parameters $D$ (depth), $C$ (channels), and $K$ (kernel size) on multi-label datasets Physionet2021 and Alibaba. Through extensive experiments, it finds that shallower networks with wider channels and smaller kernels generally improve performance, though the best scale is task-dependent. The study also links receptive field size and Global Average Pooling to performance, showing that ECG benefits from smaller receptive fields and that GAP acts as a regularizer when the field is local. The findings offer practical guidance for designing efficient, accurate ECG classifiers with reduced compute and training time, and highlight label-specific differences that suggest tailored architectures for individual clinical tasks.

Abstract

We study scaling convolutional neural networks (CNNs), specifically targeting Residual neural networks (ResNet), for analyzing electrocardiograms (ECGs). Although ECG signals are time-series data, CNN-based models have been shown to outperform other neural networks with different architectures in ECG analysis. However, most previous studies in ECG analysis have overlooked the importance of network scaling optimization, which significantly improves performance. We explored and demonstrated an efficient approach to scale ResNet by examining the effects of crucial parameters, including layer depth, the number of channels, and the convolution kernel size. Through extensive experiments, we found that a shallower network, a larger number of channels, and smaller kernel sizes result in better performance for ECG classifications. The optimal network scale might differ depending on the target task, but our findings provide insight into obtaining more efficient and accurate models with fewer computing resources or less time. In practice, we demonstrate that a narrower search space based on our findings leads to higher performance.

Figures (12)

• Figure 1: Classification performance of networks with different layer depth ($D$), the number of channels ($C$), and kernel size ($K$) depending on database sources.
• Figure 2: The classification performance of networks with different layer depth, number of channels, and kernel size on arrhythmia labels in Physionet 2021 dataset.
• Figure 3: The classification performance of networks with different layer depth, number of channels, and kernel size on conduction disorder labels in Physionet 2021 dataset.
• Figure 4: The classification performance of networks with different layer depth, number of channels, and kernel size on axis deviations labels in Physionet 2021 dataset.
• Figure 5: The classification performance of networks with different layer depth, number of channels, and kernel size on prolonged intervals labels in Physionet 2021 dataset.