Sleep structure discriminates patients with isolated REM sleep behavior disorder: a deep learning approach.

Abstract

Rapid eye movement (REM) sleep behavior disorder (RBD) is a disorder characterized by increased muscle tone and dream-enactment behaviors in REM sleep. In its isolated form (iRBD), it is a prodromal stage of neurodegenerative diseases. Currently, diagnosis of RBD requires time-consuming and subjective visual inspection of polysomnography (PSG). We propose a novel fast and objective deep learning model to identify patients with iRBD based on their sleep structure. A total of 86 iRBD and 81 controls, who underwent PSG, were included in the study. A validated algorithm was used to generate hypnodensity graphs (i.e., probabilistic representations of sleep structure). A ResNet-18 model was trained on five datasets consisting of whole night hypnodensities (with and without augmentation), and shorter segments (4 hours, 2 hours, and 30 minutes) to discriminate iRBD from controls. Using entire-night hypnodensity had notable benefits in terms of performance compared to shorter length segments, leading to a mean macro F1 score of 0.717 (per-segment), and of 0.784 (per-subject). Our findings show that sleep structure is important for iRBD classification and could potentially help clinicians.