Improving ECG Classification through Neural Network Scale Optimization

[Submitted on 24 Aug 2023]

Download a PDF of the paper titled “Optimizing Neural Network Scale for ECG Classification” by Byeong Tak Lee and 2 other authors: Download PDF

Abstract: We investigate the scaling of convolutional neural networks (CNNs), specifically focusing on Residual neural networks (ResNet), for analyzing electrocardiograms (ECGs). While ECG signals are time-series data, CNN-based models have shown superior performance compared to other neural network architectures in ECG analysis. However, previous studies in ECG analysis have often overlooked the importance of network scaling optimization, which significantly enhances performance. We explore and demonstrate an efficient approach to scale ResNet by examining the effects of critical parameters such as layer depth, number of channels, and convolution kernel size. Through extensive experiments, we discover that a shallower network, a larger number of channels, and smaller kernel sizes result in improved performance for ECG classification. The optimal network scale may vary depending on the specific task, but our findings provide valuable insights for achieving more efficient and accurate models with fewer computational resources and less time. In practice, we demonstrate that a narrower search space based on our findings leads to higher performance.