IGEA: a Framework for Sleep Tracking via Edge-based Federated Learning on Multimodal Data

Monitoring sleep quality is crucial for personalized health and cognitive optimization, yet wearable devices fail to consider the influence of environmental factors. Integrating this data often requires centralized aggregation, which can raise privacy concerns. To address this, we introduce IGEA, a federated learning (FL) framework designed for smart sleep monitoring that ensures privacy preservation. IGEA combines wearable-derived physiological data with ambient data collected from Raspberry Pi 4 nodes, each functioning as an autonomous FL client. Given the absence of publicly available datasets that combine both physiological and environmental data, IGEA employs a two-stage pipeline. First, a model pre-trained on public wearable datasets estimates sleep quality, followed by a neural network (NN) refined through FL on local ambient data to improve the prediction. We implemented IGEA on real clients and evaluated it under varying data loads. Results demonstrate that FL significantly outperforms local-only training, achieving near-centralized model accuracy while maintaining privacy. This showcases the potential of edge FL for scalable, privacy-preserving digital wellness solutions in shared environments such as dormitories and smart homes.