Explainable Sleep Staging Algorithm using a Single-channel Electroencephalogram

Abstract

The evaluation of sleep stages is the most crucial part in diagnosing and treating patients with sleeping disorders. However, in most healthcare environments, doctors evaluate sleep stages manually by using patients’ polysomnography (PSG) data, which leads to high economic and time costs. PSG data are extremely complicated due to the amount of data and its recording process. In this study, instead of using PSG data single-channel EEG data are used to create an automated model for evaluating the five stages of sleep. The proposed model is an explainable artificial intelligence model for applications in a real-world medical environment. For this purpose, single-channel EEG data are decomposed into each signal component by band-pass filters. For post-hoc analysis, the learning rate for each key component in determining the sleep stages was estimated using the attention mechanism. A cross-evaluation was conducted on data from 80 subjects. The result was an averaged F1-score of 72.66 (±22.24) and an explainable model where EEG components were more effective in estimating each sleep stage.