Federated Learning for Sleep Stage Classification on Edge Devices via a Model-Agnostic Meta-Learning-Based Pre-Trained Model

Sleep stage classification using physiological data obtained from consumer wearable devices has gained significant attention. However, data privacy regulations prevent the collection of individual data for training the models. Federated learning (FL) enables collaborative training across distributed networks of edge devices without sharing private data. This paper presents a federated learning approach that leverages computational resources of Internet of Thing (IoT) edge devices for sleep stage classification. Given the limited computational power and resources on edge devices, we evaluate the impact of model initialization in FL using pre-trained weights of model-agnostic meta-learning (MAML) models. We evaluate the performance of on-device sleep stage classification by comparing local models to different FL settings initialized with random weights (FL-Random) and pre-trained weights (FL-ML). FL-ML enables faster model convergence, yielding better classification performance (up to 22.6%) in a small number of training rounds, compared to local models. Starting from FL-ML reduces the training time required to reach a target accuracy and allows for the training of more accurate models, compared to initializing FL from random weights or using local models. Our study highlights the feasibility of FL initialized with pre-trained weights in real-world IoT settings for sleep stage classification on edge devices.