ScholarWorks@숙명여자대학교

초록

Wearable healthcare consumer electronics generate substantial medical time-series data with significant potential for personalized healthcare applications. However, effectively modeling such data in federated settings presents unique challenges due to pronounced Non-Independent and Identically Distributed (Non-IID) characteristics, privacy concerns, and personalization requirements. This paper proposes Med-FedLSG, a novel federated learning framework that addresses these challenges through three key innovations: a variational temporal representation learning mechanism with explicit disentanglement of shared and personalized features, a conditional temporal generator with physiological constraints, and a two-level federated optimization framework based on knowledge distillation. Extensive experiments on the MIMIC-III clinical database and UCI-HAR dataset demonstrate that Med-FedLSG consistently outperforms existing federated learning methods, achieving 85.47% accuracy on MIMIC-III and 92.28% on UCI-HAR. Furthermore, our framework achieves superior intra-user consistency scores of 0.89 on MIMIC-III and 0.93 on UCI-HAR, demonstrating enhanced stability for long-term medical monitoring. Ablation studies validate the effectiveness of our personalized representation separation mechanism and conditional temporal generator. The proposed approach successfully balances generalization and personalization while maintaining privacy, offering promising solutions for smart wearable health applications.

키워드