Deep-Learning-Based Automated REM Sleep Detection in Patients With REM Sleep Behavior Disorder: Is It Reliable?

Abstract

Background and purpose: Rapid eye movement (REM) sleep without atonia makes it difficult to detect REM sleep stages using electromyography in patients with REM sleep behavior disorder (RBD). The objectives of this study were to propose an automated REM sleep detector that requires only electroencephalography (EEG) and electrooculography (EOG) data, and to evaluate its performance using real-world polysomnography (PSG) data in RBD patients.

Methods: This multicenter study used 310 PSG datasets obtained from 5 tertiary hospitals. The data were divided into RBD (n=200) and non-RBD (n=110), as well as, into Parkinson's disease (PD) with RBD (n=76), PD without RBD (n=46), idiopathic RBD (iRBD) (n=124), and healthy controls (n=64). An automated computerized REM detection algorithm was implemented using U-Sleep's publicly available pretrained network.

Results: The U-Sleep-based REM sleep-detection algorithm correctly identified REM sleep with an area under the receiver operating characteristic curve (AUC) of 0.90±0.14. The classification performance of the REM sleep detector differed significantly between RBD and non-RBD patients (AUC=0.88±0.13 vs. 0.93±0.14, p=0.007). The REM sleep detector accurately classified REM sleep in the order of healthy controls, PD without RBD, iRBD, and PD with RBD, with AUC values of 0.94±0.02, 0.92±0.03, 0.90±0.02, and 0.86±0.02, respectively.

Conclusions: Our U-Sleep-based REM sleep detector based on only EEG and EOG data showed good performance in detecting REM sleep. However, it performed considerably worse in RBD, especially in PD with RBD. Using transfer learning with fine-tuning by expert review, a high-performance REM sleep-detecting system will be realized.

Keywords: REM sleep behavior disorder; REM sleep detector; REM sleep without atonia; automated algorithm.

Fig. 1. Overview of the sleep staging comparison process using U-Sleep and expert annotations. PSG data from two EOG channels (left and right) and six EEG channels (F3, F4, C3, C4, O1, and O2) were processed. A single EEG–EOG pair was randomly selected, and patient information was anonymized before sending the data to the U-Sleep API, which returned hypnograms. This process was repeated for all 12 combinations, averaging the results. Expert staging used all channels, and the outputs were compared, focusing on the accuracy of REM versus NREM. DOB, date of birth; EEG, electroencephalography; EOG, electrooculography; N1, NREM stage 1; N2, NREM stage 2; NREM, non-REM; REM, rapid eye movement.

Fig. 2. Performance of the U-Sleep-based REM sleep detector for RBD and non-RBD. A: ROC curves and AUCs for the non-RBD (blue) and RBD (red) groups. B: Accuracy (pink), sensitivity (blue), and specificity (green) of each test set in the non-RBD and RBD groups. AUC, area under the receiver operating characteristic curve; RBD, REM sleep behavior disorder; REM, rapid eye movement; ROC, receiver operating characteristic.

Fig. 3. False-positive sleep-stage distribution of each test fold in the non-RBD (A) and RBD (B) groups. N1, N2, and W correspond to pink, green, and blue, respectively. N1, NREM stage 1; N2, NREM stage 2; NREM, non-REM; RBD, REM sleep behavior disorder; REM, rapid eye movement; W, wake.

Fig. 4. Performance of the U-Sleep-based REM sleep detector (subgroup analysis). A: AUC in group 1 (PD with RBD, pink), group 2 (PD without RBD, purple), group 3 (idiopathic RBD, red), and group 4 (healthy control, mint). Error bars represent standard errors of the mean. B: Accuracy (pink), sensitivity (blue), and specificity (green) of each test fold in the four subgroups. AUC, area under the curve; RBD, REM sleep behavior disorder; REM, rapid eye movement; PD, Parkinson’s disease.