A Deep Learning Approach for Classifying Rapid Eye Movement Sleep Behavior Disorder Using EEGNet.

BACKGROUND AND PURPOSE: Rapid eye movement (REM) sleep behavior disorder (RBD) is a prodromal marker for α-synucleinopathies, and its early and accurate diagnosis provides invaluable insights into the development of neurodegenerative diseases. This study aimed to propose a deep-learning-based automated framework for RBD classification and to evaluate its performance using real-world, multi-cohort polysomnography (PSG) data.

METHODS

REM electromyogram (EMG) data of 227 PSG recordings collected across five tertiary hospitals were analyzed. The dataset was divided into RBD (n=156) and non-RBD (n=71) and further stratified into Parkinson's disease (PD) with RBD (n=60), PD without RBD (n=24), isolated RBD (n=96), and healthy controls (n=47). An automated framework for RBD classification, based on manual and automated sleep staging, was implemented using EEGNet.

RESULTS

The EEGNet-based classifier achieved a receiver operating characteristics-area under the curve of 0.812 (95% confidence interval [CI], 0.744-0.871) using manual staging, which was significantly higher than the 0.777 (95% CI, 0.704-0.843) achieved with the automated U-Sleep approach (p=0.039). However, relying solely on REM EMG data proved insufficient to accurately distinguish among the four clinical subgroups.

CONCLUSIONS

Our EEGNet-based RBD classifier, focused solely on EMG data, demonstrated reasonable performance in detecting RBD. This proposed model has the potential to reduce clinicians' workload and support diagnostic decision-making and enabling more accurate RBD classification. This multi-center study highlights the feasibility of an automated and generalizable RBD detection framework, paving the way for broader clinical implementation.