Category: Other
Objective: Evaluate the performance of a classifier using computer vision trained on video-polysomnography (vPSG) data recorded in the sleep laboratory setting over a single night using conventional equipment.
Background: Isolated Rapid Eye Movement (REM) Sleep Behavior Disorder (iRBD) is the strongest predictor of Parkinson’s disease and its diagnosis requires vPSG. However, interpreting vPSG can be challenging and incidental RBD cases are easily missed. Prior work using 3D cameras showed proof of feasibility and high performance of a machine learning-trained classifier using computer vision, reaching an accuracy of 0.86.
Method: We analyzed 2D video data from vPSG recordings collected at the Stanford Sleep Center. The set included 78 patients with definite iRBD (66.5 age mean, and 80.2% males) and 109 individuals without RBD (64.1 age mean and 73.3% males, including 41 obstructive sleep apnea; 37 restless legs/periodic limb movements, 9 non-RBD parasomnias, 5 insomnia, 2 narcolepsy, and 39 with normal sleep).
We used an automatic computer vision algorithm, i.e., optical flow, to analyze video data and identify periods of movement during REM sleep. The 4 movement-related extracted features were: rate (frequency of movements), ratio (proportion of REM sleep with movements), intensity (area of moving parts) and velocity (rate of change of movements).
We utilized a logistic regression classifier trained and tested using a 10-fold cross-validation scheme, exploring various combinations of the aforementioned features.
Results: The average number of REM movement periods were 58.13 and 42.53 in iRBD and controls, respectively. Best accuracy and F1 scores were 0.86 and 0.85, respectively. Adding total apnea-hypopnea (all night, REM and non-REM) and periodic limb movement index improved the sensitivity (0.87) and negative predictive value (0.88). Adding age and gender as input features increased sensitivity to 0.91 and F1 score to 0.87.
Conclusion: Demographic characteristics and automatically extracted features related to movements during REM sleep allow to classify iRBD versus other sleep disorders and normal sleep. Using conventional 2D cameras currently used in all sleep labs, this approach could assist expert scorers in interpreting vPSG and diagnosing iRBD.
To cite this abstract in AMA style:
M. Abdelfattah, O. Sum-Ping, J. Galati, S. Marwaha, A. Alahi, E. During. Enhancing iRBD diagnosis through 2D video analysis: a machine learning approach [abstract]. Mov Disord. 2024; 39 (suppl 1). https://www.mdsabstracts.org/abstract/enhancing-irbd-diagnosis-through-2d-video-analysis-a-machine-learning-approach/. Accessed October 12, 2024.« Back to 2024 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/enhancing-irbd-diagnosis-through-2d-video-analysis-a-machine-learning-approach/