Comparison of Federated Learning Strategies on ECG Classification

Federated Learning has garnered considerable attention in recent years due to its capability to maintain data at its original location, thus preserving privacy and security while still yielding a high-quality generalized model. Like the privacy of any data, it is of paramount importance to protect the privacy of health data such as ECG recordings. Federated Learning, in this regard, can offer privacy and security in the classification of ECG data. In our study, we drew comparisons among the centralized model, Federated Learning models trained with both Independently and Identically Distributed (IID) data, and those trained with Non-IID data. For each model, we used Convolutional Neural Network (CNN) architecture, utilizing the MIT-BIH Arrhythmia Database as our dataset. While we achieve 98.39% accuracy with the centralized model, FedAvg with IID and Non-IID data provides 98.53%, and 87.70%, respectively. The results underline the effectiveness and challenges of various training data types in Federated Learning for ECG classification task.