Deep learning in the cross-time frequency domain for sleep staging from a single-lead electrocardiogram

This study classifies sleep stages from a single lead electrocardiogram (ECG) using beat detection, cardiorespiratory coupling in the time-frequency domain and a deep convolutional neural network (CNN).

An ECG-derived respiration (EDR) signal and synchronous beat-to-beat heart rate variability (HRV) time series were derived from the ECG using previously described robust algorithms. A measure of cardiorespiratory coupling (CRC) was extracted by calculating the coherence and cross-spectrogram of the EDR and HRV signal in 5 min windows. A CNN was then trained to classify the sleep stages (wake, rapid-eye-movement (REM) sleep, non-REM (NREM) light sleep and NREM deep sleep) from the corresponding CRC spectrograms. A support vector machine was then used to combine the output of CNN with the other features derived from the ECG, including phase-rectified signal averaging (PRSA), sample entropy, as well as standard spectral and temporal HRV measures. The MIT-BIH Polysomnographic Database (SLPDB), the PhysioNet/Computing in Cardiology Challenge 2018 database (CinC2018) and the Sleep Heart Health Study (SHHS) database, all expert-annotated for sleep stages, were used to train and validate the algorithm.

Ten-fold cross validation results showed that the proposed algorithm achieved an accuracy (Acc) of 75.4% and a Cohen's kappa coefficient of [Formula: see text]  =  0.54 on the out of sample validation data in the classification of Wake, REM, NREM light and deep sleep in SLPDB. This rose to Acc  =  81.6% and [Formula: see text]  =  0.63 for the classification of Wake, REM sleep and NREM sleep and Acc  =  85.1% and [Formula: see text]  =  0.68 for the classification of NREM sleep versus REM/wakefulness in SLPDB.

The proposed ECG-based sleep stage classification approach that represents the highest reported results on non-electroencephalographic data and uses datasets over ten times larger than those in previous studies. By using a state-of-the-art QRS detector and deep learning model, the system does not require human annotation and can therefore be scaled for mass analysis.