A Lightweight Split Federated Learning Approach With Dynamic Model Aggregation for Breast Cancer Prediction in Consumer-Centric Health Electronics

With disease screening services increasingly becoming accessible to individual users, it is of utmost importance to ensure accurate and efficient risk prediction while protecting sensitive medical information in consumer-oriented health electronics. Although deep learning shows great potential in medical image analysis, centralized methodologies encounter hurdles, including high privacy risks, substantial computational demands, and limited scalability. Federated learning addresses these limitations by enabling collaborative model training without the need to exchange raw data. However, it still suffers from client heterogeneity, inadequate communication efficiency, and sluggish convergence. To address these challenges, we propose a novel Split Federated Learning (SFL) approach enhanced by a Lightweight Dynamic Model Aggregation strategy (named as LDA-SFL), tailored explicitly for resource-constrained and privacy-sensitive healthcare scenarios. Our method leverages SFL to effectively balance computational loads across devices, improve communication efficiency, and strengthen privacy protection. In addition, we design a client selection mechanism that dynamically adjusts the aggregation strategy by filtering out models that may degrade global performance, thereby improving training efficiency and overall performance. Extensive experiments demonstrate the effectiveness of LDA-SFL in breast cancer prediction. Compared to conventional distributed learning methods, LDA-SFL significantly reduces aggregation overhead and improves overall training efficiency.