Distributed data-privacy preserving federated learning method for sleep stage classification

Abstract

Artificial intelligence and big data have great promise for promoting sleep medicine assistance. However, obstacles such as data privacy restrictions occur when enough data are collected from different data centers to develop classification models with better performance. Federated learning (FL), a collaborative machine learning approach, aims to address problems by acquiring models rather than directly accessing data from multiple clients. In this work, we utilize federated learning to increase model performance and preserve data privacy when deploying automatic sleep staging methods with insufficient data. And we construct a CNN-Transformer model and design an anti-class ratio (ACR) weighted loss function to complete the classification task with class imbalanced data. We evaluate our model and ACR using Sleep-EDF database and achieve better overall and N1 classification performance. To clearly show the efficiency of FL, we set three different data volume experiments with three independent heterogeneous public sleep databases. The experimental results are compared with two benchmark methods, traditional local learning and data-centralized learning. The results suggest that federated learning method outperforms the local learning approach in classification performance and achieves comparable results with data-centralized learning. As a result, we argue that federated learning is a promising alternative for multiple clients developing sleep staging algorithms especially with insufficient data because it adds benefits in terms of model performance and data privacy protection.