0689 Automated Detection of Isolated REM Sleep Behaviour Disorder Using Computer Vision
Mohamed Abdelfattah,Oliver Sum-Ping,Joanna Galati,Salonee Marwaha,Alexandre Alahi,Emmanuel During
DOI: https://doi.org/10.1093/sleep/zsae067.0689
IF: 6.313
2024-04-20
SLEEP
Abstract:Abstract Introduction Isolated Rapid Eye Movement (REM) Sleep Behavior Disorder (iRBD) is a marker of neurodegeneration and is currently diagnosed based on REM sleep without atonia (RSWA) and a history of dream enactment. However, even for sleep experts, scoring RSWA can be challenging and incidental RBD cases are easily missed. Previous research focused on the automatic identification of iRBD using 3D video analysis. This study aims at developing a machine learning classifier to detect iRBD using only conventional 2D cameras and evaluating its performance on an expanded cohort. Methods We used 2D video data from in-lab video-polysomnography recorded at the Stanford Sleep Center in 78 patients with definite iRBD per ICSD-3 TR criteria and 109 without RBD (41 OSA [AHI > 15], 37 RLS/PLM [PLMi>15], 9 non-RBD parasomnias, 5 insomnia, 2 narcolepsy, 39 with normal sleep). An automatic computer vision algorithm, i.e. optical flow, was applied on the recorded videos to detect periods of movements in REM sleep. Movement periods were divided based on their duration from very brief (0.1s) to extended (>30s) motor behaviors, 1 second of immobility separating each period. Features were movement frequency and proportion of movement periods in REM sleep. Different combinations of these features fed a logistic regression classifier, which was trained and tested in the 10-fold cross-validation scheme. Results On average, 58.13 and 42.53 movement periods were detected during REM sleep in cases and controls, respectively. The best performance was achieved using short movements (0.1-2s) with sensitivity of 0.921, specificity of 0.674, accuracy of 0.835 and F1 score of 0.847. With medium (2-15s) and long (15-300s) movements, accuracy was 0.845 and 0.672, sensitivity 0.743 and 0.147, specificity 0.901 and 1.0, and F1 score 0.808 and 0.256, respectively. Adding features of gender, AHI, and PLM index to short movements increased accuracy to 0.886. Conclusion One-night video data from conventional video-polysomnography can detect iRDB using machine learning techniques. Automated video analysis of REM sleep could augment our current ability to screen and diagnose RBD in the sleep laboratory setting. Support (if any)
neurosciences,clinical neurology