Parkinson’s disease, a neurodegenerative disorder that affects more than 10 million people worldwide, is traditionally diagnosed clinically by finding slow movement, rest tremors (i.e., a shaking of limbs), and muscle rigidity. But spotting it early isn’t easy — no specific, objective diagnostic test for Parkinson’s exists.
Encouragingly, scientists at the University of Oxford have made headway in developing a framework they claim can automatically detect an early predictor of Parkinson’s: rapid eye movement (REM) sleep behavior disorder (RBD). They describe it in a new paper (“Detection of REM Sleep Behaviour Disorder by Automated Polysomnography Analysis“) published on the preprint server Arxiv.org.
“There is clear evidence that RBD is a precursor to Parkinson’s disease, Lewy Body disease, and multiple system atrophy, preceding them by years,” the researchers wrote. “Therefore, an accurate RBD diagnosis would provide invaluable early detection and insights into the development of these neurodegenerative disorders … In this study, we propose a fully automated pipeline for RBD detection.”
A few automated scoring algorithms for RBD already exist, they noted, which take into account polysomnography and evidence of REM sleep without atonia — the two requirements for an RBD diagnosis as standardized by the International Classification of Sleep Disorders. But not many of them are designed for older individuals, or for those who suffer from sleep disorders.
It’s a fundamentally different approach than that taken by researchers at the Institute for Robotics and Intelligent Systems in Zurich, Switzerland, who in a paper published in October detailed an AI system that can diagnose Parkinson’s with data collected from a suite of smartphone-based tests.
Designing and testing the model
In building a dataset, the Oxford scientists sourced sleep study records from 53 patients from the Montreal Archive of Sleep Studies, an open access database of laboratory-based recordings. All were annotated by an expert and preprocessed to reduce noise.
To classify each sleep stage, the researchers used a random forest (RF) model — a supervised learning algorithm that constructs an ensemble of decision trees and outputs the mean prediction of the individual trees — and 156 features extracted from electroencephalograms (records of brain activity), electrooculograms (records of eye movements), and electromyograms (records of electrical activity produced by skeletal muscles) sourced from the sleep study notes.
For RBD detection, the RF classifier was trained with techniques to quantify muscle atonia (a condition in which a muscle loses its strength) and additional features. (Atonia turned out to be the most important predictor of RBD.) In testing, accuracy improved by 10 percent to 96 percent when using manually annotated sleep staging and remained high (92 percent) when using automated sleep staging.
The team noted that the results could be further improved with better automated sleep stage classification — potentially a technique involving deep learning, layered mathematical functions that mimic the behavior of neurons in the brain.
And they said that future work will investigate how the RBD detection framework applies to the clinical environment and how non-motor features — such as altered heart rate variability during sleep — might help delineate RBD. The team will also look to apply improved sleep staging algorithms across a range of disorders while incorporating signals to enhance RBD detection.
“The algorithm outperforms individual metrics,” the researchers write. “[The] study validates a tractable, fully-automated, and sensitive pipeline for RBD identification that could be translated to wearable take-home technology,”