Introduction: Annotating sleep and sleep events is a critical step in the analysis of sleep data. Manual annotation is a tedious and time-consuming task and requires expertise. Studies have shown that computer-assisted annotation can improve the time required and the quality of the annotation. In this study, we evaluate the performance of a single deep learning model for detecting sleep and sleep events.

Method: We developed a convolutional u-net model for the detection of sleep, arousal, respiratory and movement events. The model was trained on 2291 and tested on 1202 overnight polysomnograms (PSG) recorded at the University Hospital Carl Gustav Carus in Dresden, Germany. We used electroencephalogram (EEG), electrooculogram (EOG), electromyography (EMG), snoring, flow, body position, effort and blood oxygen channels from the PSG. The signals are downsampled to 50 Hz and the model produces a 50 Hz probability of binary detection for sleep, arousal, respiratory event and periodic leg movement (PLM) as output. We optimised the threshold for each component and combined consecutive segments into single events. We used F1 scores to compare the detected events with any overlapping real events and calculated several sleep parameters (e.g. total sleep time (TST), arousal index (ArI), apnoea-hypopnoea index (AHI) and periodic limb movement index (PLMNi)) and compared the results of automatic detection with the manual annotations.

Results: We achieved F1 values of 0.92, 0.71, 0.75 and 0.80 for sleep, arousal, respiratory event and PLM. For the sleep parameters, we obtained a correlation of 0.89 for TST, 0.85 for ArI, 0.90 for AHI and 0.97 for PLMNi (index calculated for true TST). Using the calculated TST, we obtained a correlation of 0.65 for ArI, 0.86 for AHI and 0.77 for PLMi.

Conclusion: Actual implementation in the clinical setting is challenging and the error is increased when dependent sleep parameters are calculated. Still, our model achieves state of the art results using data from daily use in a sleep laboratory. The use of multiple events shows no negative impact on detection compared to single event prediction. This approach demonstrates that a full evaluation of sleep with a single model is possible in a clinical setting.