Multi-Cohort Evaluation of an Automated Sleep Stage Detection Methodology Using ECG and Respiration Signals *

This study presents an automated sleep stage detection methodology using a reduced number of signals, as input from polysomnography (PSG) recordings. The aim is to establish a competitive sleep stage detection performance based on AI models for respiration and heart rate signal analytics, as these signals can be effectively collected by wearable and wireless monitoring solutions in home and clinical environment. A wide range of time, frequency and time-frequency domain features were first extracted in 30-sec long signal segments, along with heart and respiration rate variability analytics. The most optimal subset of features per evaluation run was assessed and selected using mutual information and each segment was then classified as either Wake, N1+N2, N3 or REM class, using a Gradient Boosted Decision Tree classifier. The proposed methodology was evaluated with data from two different databases, containing both healthy subjects and patients with apnea-related disorders, achieving an average classification accuracy of 84.62% and 85.18%, respectively, in the challenging 4-class task of wake-light-deep-REM sleep stage detection, outperforming previous results. Reducing the model’s input to only respiration and heart rate data, the proposed methodology paves the way to the use of wireless and contactless systems, enabling prolonged and unobtrusive monitoring of patients with various sleep disorders with high sleep stage detection accuracy.