Unmasking the Prodrome: A Deep-Learning Bridge to Precision Diagnostics in Rapid Eye Movement Sleep Behavior Disorder

Rapid eye movement (REM) sleep behavior disorder (RBD) has a pivotal position in the evolution of neurodegenerative diseases. Polysomnography (PSG)-confirmed RBD is no longer considered a simple, isolated parasomnia; rather, it is firmly established as a definitive prodromal marker for α-synucleinopathies, having an exceptionally high risk of conversion to Parkinson’s disease (PD) and dementia with Lewy bodies.¹,² Consequently, the ability to accurately diagnose RBD would provide a critical opportunity beyond the management of nocturnal sleep behaviors, by enabling early interventions and the application of disease-modifying therapies before the onset of irreversible central dopaminergic neurodegeneration.

Diagnosing RBD is challenging due to substantial methodological limitations. Although detecting REM sleep without atonia depends on objective PSG data, their interpretation remains largely subjective, being highly dependent on the expertise of sleep specialists, and this is also very labor-intensive.³,⁴ This issue of the Journal of Clinical Neurology includes a report on an important methodological advancement with the validation of a deep-learning–based automated framework that accurately classifies RBD using real-world, multicenter electromyography (EMG) data.⁵

A key strength of the reported study lies in its shift away from complex, manual feature extraction toward a deep-learning approach capable of inherently capturing hidden temporal dynamics. Adopting EEGNet—a convolutional neural network highly adept at identifying spatiotemporal patterns—has enabled the successful automation of RBD detection based solely on processing EMG amplitude time-series data obtained during REM sleep. More importantly, the study has clarified the underlying decision-making process of the artificial intelligence model through the application of SHAP (Shapley Additive exPlanations) values. This confirms that the network’s predictions are explicitly driven by physiologically meaningful high-amplitude phasic muscle bursts, demonstrating that the model effectively incorporates clinical reasoning while operating at computational speeds.

The study has also directly addressed one of the most-persistent hurdles in big-data sleep medicine: interinstitutional variance. Baseline EMG amplitudes vary substantially between hospitals due to differences in equipment and impedance levels. Robustly implementing an intrasubject normalization strategy—scaling REM-sleep EMG amplitudes against each individual’s non-REM (NREM) baseline—has successfully harmonized data from multiple distinct tertiary-care centers. This approach highlights the importance of developing scalable and generalizable algorithms capable of handling the heterogeneity inherent in multicenter clinical environments.

However, critical challenges remain to be addressed. Although the framework achieved impressive diagnostic performance relative to expert manual sleep staging, its accuracy and specificity decreased significantly when coupled with an automated sleep staging system. In patients with fragmented sleep architectures and abnormally elevated muscle activity during REM sleep owing to neurodegeneration, the automated system struggled to accurately demarcate the boundaries between REM sleep, wakefulness, and NREM sleep. The resulting incorporation of misclassified epochs in the input data increased the rate of false-positive predictions. Furthermore, the routine clinical implementation of this technology is currently constrained by its strict reliance on high-quality, artifact-free input signals and the requirement for a dedicated computational infrastructure. Extensive prospective validation in diverse, real-world cohorts remains necessary to firmly establish its diagnostic reliability before it can be widely deployed.

The technological advancements required to overcome these limitations will directly translate into widespread clinical innovations. Foremost among these is an expansion in diagnostic accessibility. By fusing multimodal data such as electroencephalography and electrooculography and refining the performance through robust domain adaptation techniques, future iterations of this deep-learning algorithm could substantially lower the barriers to its implementation in diagnostic environments. Large-scale, population-level screening for RBD could become a reality, replacing cost-prohibitive in-laboratory PSG with the utilization of the restricted biosignals acquired from home sleep apnea testing and commercial wearable devices.⁶

Beyond diagnostic screening, this technology could become a pivotal patient stratification tool for use in precision medicine and novel drug development. If deep-learning algorithms can improve beyond binary disease classification to track subtle, quantitative longitudinal changes in REM-sleep muscle activity, they could act as sophisticated digital biomarkers for predicting the precise window of phenoconversion from isolated RBD to PD or dementia with Lewy bodies.⁷ Consequently, this approach will serve as an essential infrastructure in future clinical trials for α-synuclein–targeted agents and disease-modifying therapies, accurately identifying patients during the optimal window for intervention and thereby maximize therapeutic efficacy.

While recent advances in neuroimaging and fluid biomarkers have revealed the biological trajectories of Lewy-body diseases, the automated algorithm presented in the reported study provides a practical mechanism to implement these insights at scale in real-world clinical practice. This reported deep-learning–based biosignal analysis framework offers a valuable approach to alleviating the interpretative burden on clinicians. Furthermore, it provides a methodological foundation that could potentially support future early intervention strategies prior to the onset of overt clinical symptoms.

Availability of Data and Material

Data sharing not applicable to this article as no datasets were generated or analyzed during the study.