Self-Evolving Federated Learning Framework for Real-Time Personalized Disease Risk Prediction Using Multimodal Wearable and Genomic Data

Personalized disease risk prediction plays a crucial role in enabling proactive and preventive healthcare systems. However, achieving accurate and timely predictions remains challenging due to the presence of high-dimensional multimodal data and increasing concerns regarding data privacy. Wearable devices continuously generate large volumes of physiological signals such as heart rate, activity level, and sleep patterns, while genomic profiles provide static yet high-resolution biological information. Effectively integrating these heterogeneous data sources in a secure manner is non-trivial. To address these challenges, this work proposes a SELF-Evolving Federated Learning (SELF-FL) framework for real-time personalized disease risk prediction. The proposed framework adopts decentralized model training, ensuring that sensitive user data remains on local devices and is never directly shared. SELF-FL incorporates neural architecture search (NAS) to enable self-evolving model adaptation, allowing the learning architecture to dynamically adjust based on data characteristics. Additionally, cross-modal fusion mechanisms are employed to effectively combine wearable-derived temporal features with genomic embeddings. Federated aggregation is used to update a global model while preserving data privacy. Experimental evaluations conducted on both synthetic and real-world multimodal datasets indicate that SELF-FL consistently outperforms conventional centralized learning and standard federated approaches. The framework demonstrates improved predictive accuracy, robustness to data heterogeneity, adaptability to evolving inputs, and better interpretability, making it suitable for scalable and privacy-compliant personalized healthcare analytics.