Efficient ECG Classification with Light Weight Shuffle GhostNet Architecture

The electrocardiogram (ECG) signal is critical in diagnosing cardiovascular disorders. Deep learning-based ECG classification has recently attracted attention to improving the performance of heart disease treatment. Most prior classification algorithms focused on increasing accuracy while ignoring the time process caused by a heavy model. Improving speed with lightened-weight design is critical for reducing time and expense in real-world applications. To address this issue, this paper suggests a novel strategy for building a lightweight network by employing the GhostNet Architecture without sacrificing accuracy. The residual ShuffleGhostBottleneck block is constructed based on the cheap operation and various modules such as ChannelShuffleModule, GhostModule, and Efficient Channel Attention module. Our strategy assists the network in obtaining more features while reducing parameters. The PhysioNet/Computing in Cardiology Challenge 2021 evaluates our proposal on 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead. The suggested method produced non-official challenge scores of 0.58, 0.55, 0.57, 0.56, and 0.56 on 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead ECG test sets. Compared to the same testing on another architecture, the total parameters are reduced by about 7.6 times, and GFLOPS decreased by 20 times. The proposed models have a clinical applicability potential.