Federated Learning for Arrhythmia Detection of Non-IID ECG

In this paper, a distributed arrhythmia detection algorithm based on electrocardiogram (ECG) is proposed for auxiliary diagnosis and treatment. ECG that contains tremendous cardiac rhythm information plays an important role in clinical treatment. Machine learning (ML) algorithms can effectively build the relationship between ECG and the underlying arrhythmia in it. Due to the privacy sensitivity of the ECG, we introduced a federated learning (FL)-based distributed algorithm that enables each medical institution to cooperatively train a arrhythmia detection algorithm locally. Compared with the traditional centralized ML algorithms, the use of FL-based algorithm does not need to collect all the local ECG of each medical institution to an external platform to perform centralized learning, and hence preventing the privacy from leakage. However, ECG collected from different medical institution is non-independent and identically distributed (non-IID) in reality, which will lead to non convergence of the FL-based algorithm. To address this challenge, we optimize the FL-based algorithm using a sharing strategy for partial ECG data of each medical institution combined with elastic weight consolidation (EWC) algorithm. Here, the sharing strategy, which makes each medical institution share ECG data to the central server while not share to other clients, could help build an initial FL model and EWC algorithm make the accuracy of the model trained by each medical institution not decline, therefore the proposed FL algorithm can achieve a trade-off between the privacy and model performance. The experiment results show that, compared with baseline FedAvg algorithm and FedCurv algorithm, the optimized FL-based algorithm is faster in convergence for IID ECG and achieves signicant improvement in terms of both recall and precision for non-IID ECG.