AI-enhanced algorithm improves diagnosis of REM sleep behavior disorder

A Mount Sinai-led team of researchers has improved an artificial intelligence (AI)-powered algorithm to analyze video recordings of clinical sleep tests, ultimately improving the accurate diagnosis of a common sleep disorder that affects more than 80 million people worldwide. The findings of the study have been published in the journal Annals of Neurology on January 9.

REM sleep behavior disorder (RBD) is a sleep condition that causes abnormal movements or physical actions from dreams during the rapid eye movement (REM) phase of sleep. RBD that occurs in otherwise healthy adults is called “isolated” RBD. It affects more than a million people in the United States and is an early sign of Parkinson’s disease or dementia in almost all cases.

RBD is extremely difficult to diagnose because its symptoms can go unnoticed or be confused with other diseases. A definitive diagnosis requires a sleep study, known as a video polysomnogram, performed by a medical professional at a facility with sleep monitoring technology. The data is also subjective and can be difficult to interpret universally based on multiple and complex variables, including sleep stages and amount of muscle activity. Although video data is systematically recorded during a sleep test, it is rarely reviewed and often discarded after the test has been interpreted.

Previous limited research in this area had suggested that research-grade 3D cameras may be needed to detect movement during sleep, as sheets or blankets would cover the activity. This study is the first to outline the development of an automated machine learning method that analyzes video recordings routinely collected with a 2D camera during overnight sleep tests. This method also defines additional “classifiers” or features of movements, yielding an accuracy rate for detecting RBD of almost 92 percent.

This automated approach could be integrated into the clinical workflow during sleep test interpretation to improve and facilitate diagnosis and avoid missed diagnoses. This method could also be used to inform treatment decisions based on the severity of movements shown during sleep testing, ultimately helping physicians personalize care plans for individual patients.”

Emmanuel During, MD, Corresponding author, Associate Professor of Neurology (Movement Disorders) and Medicine (Pulmonary, Critical Care and Sleep Medicine), at the Icahn School of Medicine at Mount Sinai

The Mount Sinai team replicated and extended a proposal for automated machine learning analysis of movements during sleep studies created by researchers at the Medical University of Innsbruck in Austria. This approach uses computer vision, a field of artificial intelligence that allows computers to analyze and understand visual data, including images and videos. Building on this framework, experts at Mount Sinai used 2D cameras, routinely found in clinical sleep labs, to monitor patients’ sleep at night. The dataset included analysis of sleep center recordings of approximately 80 RBD patients and a control group of approximately 90 patients without RBD who had another sleep disorder or no sleep disturbance. An automated algorithm that calculated the movement of pixels between successive frames in a video was able to detect movements during REM sleep. The experts assessed the data to determine the speed, ratio, magnitude and speed of movements and the ratio of immobility. They analyzed these five characteristics of short movements to achieve the highest accuracy yet by researchers, at 92 percent.

Researchers from the Swiss Federal Institute of Technology of Lausanne (École Polytechnique Fédérale de Lausanne) in Lausanne, Switzerland, contributed to the study by sharing their expertise in computer vision.