Lightweight CNN Architectures for Real-time ECG Arrhythmia Classification on Edge Devices

Abstract

Real-time and accurate detection of cardiac arrhythmias in electrocardiogram (ECG) signals is essential in clinical decision support and remote health monitoring especially on wearable and edge devices of low power. This paper suggests a BMEL-CNN (Balanced Mixup-Enhanced Lightweight CNN) -based pipeline to classify arrhythmia based on the MIT-BIH Arrhythmia Database. The two significant issues covered by the proposed strategy include the disproportion in classes and underrepresentation of minority-classes. To begin with, there is class balancing and targeted augmentation used to balance the occurrence of N, S, V, F and Q beat types. Second, Mixup augmentation is put forward to enhance generalization by creating artificial soft-labeled samples. Lastly, a categorical focal loss, which is used together with an AdamW optimizer and cosine learning rate schedule, is utilized to emphasize learning in challenging examples and to avoid overfitting. The results of an experiment with standard MIT-BIH test split indicate that the model proposed attains 77.2% overall accuracy, macro-F1-score of 0.41 and 37.3% recall of S-class beats, which is significantly high compared to a baseline CNN model (macro-F1 = 0.37, S-class recall 0%). The network achieves similarly high recall 87.8% on V-class and 93.0% on N-class and the robustness of the network is confirmed. The last model is much smaller in size (around 220 KB) and can take around 4 ms to perform inference on a Raspberry Pi 4, which can be successfully deployed in practical real-time edge monitoring systems.