A Lightweight Model for Physiological Signals-Based Sleep Staging With Multiclass CAM for Model Explainability

IEEE Sensors Journal, vol. 24, pp. 27815–27823

Abstract

Physiological signals-based sleep staging is essential for diagnosing sleep disorders. Automatic classification using deep learning models could save time and labor, compared to manual classification and monitoring, which stimulated numerous studies aiming to improve performance with various deep learning models. However, most existing models have high computational requirements due to numerous training parameters and long input sequences (up to 15 min). Furthermore, few works provide explanations for their models’ decisions, which is crucial for reliability. To address these issues, in this article, we propose a lightweight model that can accurately classify different sleep stages with short input data and a multiclass class activation mapping (CAM) method to explain our model’s decisions. We evaluate the effectiveness of the proposed model on two public sleep datasets, namely, Sleep-EDF-20 and Sleep-EDF-78, and the method contributes an accuracy (ACC) of 86.6% and 82.8%, respectively. Additionally, our multiclass CAM method can highlight regions in the input that the model relies on for decision-making, deserving further exploration by human experts.

Authors 2

1. Yuxuan Yang

   Southwestern University of Finance and Economics

   Affiliation as printed

   School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics, Chengdu, China

2. Yanli Li Aachen Department of Radiology Division of Image Processing (LKEB)

   Leiden University Medical Center

   Affiliation as printed

   Department of Radiology, Division of Image Processing, Leiden University Medical Center (LUMC), Leiden, The Netherlands