FedECG: A federated semi-supervised learning framework for electrocardiogram abnormalities prediction

Abstract

The soaring popularity of smart devices equipped with electrocardiograms (ECG) is driving a nationwide craze for predicting heart abnormalities. Smart ECG monitoring system has achieved significant success by training machine learning models on massive amounts of user data. However, three issues arise accordingly: 1) ECG data collected from various devices may display personal characteristic variations, leading to non-independent and identically distributed (non-i.i.d.) data. These differences can impact the accuracy and reliability of data analysis and interpretation; 2) Most ECG data on smart devices is unlabeled, and data labeling is resource-consuming as it requires heavy-loaded labeling from professionals; 3) While centralizing data for machine learning can address above issues like non-i.i.d. data and labeling difficulties, it may compromise personal privacy. To tackle these three issues, we introduce a novel federated semi-supervised learning (FSSL) framework named FedECG for ECG abnormalities prediction. Specifically, we adopt a pre-processing module to better utilize the ECG data. Next, we devise a novel model based on ResNet-9 in FSSL to accurately predict abnormal signals from ECG recordings. In addition, we incorporate pseudo-labeling and data augmentation techniques to enhance our implemented semi-supervised learning. We also develop a model aggregation algorithm to improve the model convergence performance in federated learning. Finally, we conduct simulations on a real-world dataset. Experiments demonstrate that FedECG obtains 94.8% accuracy with only 50% of the data labeled. FedECG achieved slightly lower accuracy than traditional centralized methods in ECG monitoring, with a 2% reduction. In contrast, FedECG outperforms the state-of-the-art distributed methods by about 3%. Moreover, FedECG can also support unlabeled data and preserve data privacy as well.

Keywords

• Electrocardiogram
• Cardiovascular diseases
• Personal privacy
• Label scarcity
• Federated learning
• Semi-supervised learning