PFedAL-ECG: A lightweight arrhythmia diagnostic system leveraging personalized federated active learning

Electrocardiogram (ECG) is an effective non-invasive tool for detecting arrhythmia. In recent years, deep learning technologies have been widely used in ECG classification algorithms. However, data privacy, individual variability, and labeling costs have hindered the further expansion of deep learning. To tackle this challenge, this paper proposes a new arrhythmia detection algorithm called PFedAL-ECG based on personalized federated active learning. First, we build a lightweight neural network model to extract spatio-temporal features from ECG heartbeats and run it on resource-constrained edge devices. Second, we define the base and personalization layers in the federated aggregation process among the clients to build the personalized models. Third, we incorporate an improved Softmax function into the personalized model to reduce the negative aggregation effect caused by label distribution shift. Additionally, we incorporate active learning strategies into the local training of the client to reduce the cost of manual annotation. Experimental results show that the proposed method achieves satisfactory results on three public ECG databases, and cross-database experimental results prove the generalization ability of the method. Compared with other state-of-the-art methods, our method has obvious advantages in evaluation indicators, communication costs, data privacy protection, and manualannotation costs. © 2025 Elsevier B.V.