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

Area of Science:

• Neuroscience
• Sleep Medicine
• Artificial Intelligence in Healthcare

Background:

• Detecting REM sleep in REM sleep behavior disorder (RBD) is challenging due to absent muscle atonia.
• Current methods often rely on electromyography (EMG), which can be unreliable in RBD.

Purpose of the Study:

• To develop an automated REM sleep detector using only EEG and EOG data.
• To evaluate the detector's performance in patients with RBD using polysomnography (PSG) data.

Main Methods:

• Utilized 310 PSG datasets from 5 hospitals, including RBD (n=200) and non-RBD (n=110) groups.
• Employed an automated REM detection algorithm based on U-Sleep's pretrained network.
• Subdivided data into Parkinson's disease (PD) with RBD, PD without RBD, idiopathic RBD (iRBD), and healthy controls.

Main Results:

• The U-Sleep algorithm achieved an overall area under the receiver operating characteristic curve (AUC) of 0.90±0.14 for REM sleep detection.
• Performance varied significantly between RBD (AUC=0.88±0.13) and non-RBD (AUC=0.93±0.14) groups (p=0.007).
• Detection accuracy followed the order: healthy controls (0.94±0.02), PD without RBD (0.92±0.03), iRBD (0.90±0.02), and PD with RBD (0.86±0.02).

Conclusions:

• The EEG/EOG-based automated REM sleep detector demonstrates good performance.
• The system's accuracy is reduced in RBD patients, especially those with PD.
• Future improvements using transfer learning and expert fine-tuning are proposed to enhance system performance.