Identification of REM Sleep Behavior Disorder by Magnetic Resonance Imaging and Machine Learning

Jie Mei, Shady Rahayel, Christian Desrosiers, Ronald B. Postuma, Jacques Montplaisir, Julie Carrier, Oury Monchi, Johannes Frasnelli, Jean‐François GagnonmedRxiv 2021 · DOI

Abstract

Idiopathic rapid eye movement sleep behavior disorder (iRBD) is a major risk factor for synucleinopathies, such as Parkinson’s disease. This study assesses the feasibility of morphology-based machine learning (ML) in identifying and subtyping iRBD using T1-weighted structural MRI. By analyzing morphometric measurements—including thickness, surface area, volume, and deformation—the researchers trained models to classify iRBD versus controls, iRBD versus Parkinson’s disease, and iRBD with versus without mild cognitive impairment (MCI). The models achieved high accuracies (79.6% to 84.8%), identifying key discriminative regions such as the putamen, insula, and frontal gyrus. These results suggest that ML-driven neuroimaging analysis can enable efficient preclinical identification of synucleinopathies.

Introduction

Idiopathic REM Sleep Behavior Disorder (iRBD) is more than a sleep disturbance; it is a critical "prodromal" or early-stage indicator of synucleinopathies. Clinical evidence suggests that up to 80% of individuals with iRBD eventually develop Parkinson’s disease (PD) or Dementia with Lewy Bodies (DLB). However, the progression is highly heterogeneous, with some patients remaining stable for years while others decline rapidly.

Early identification is currently hindered by the high cost and limited accessibility of the gold-standard diagnostic tool, polysomnography. While structural MRI is a common clinical tool, traditional analysis often fails to capture the subtle, complex structural brain changes associated with early neurodegeneration. This study explores a multi-variate machine learning approach to leverage MRI morphometry for the detection and subtyping of iRBD.

Methodology and Data

The research team utilized a dataset of 118 participants from specialized centers in Montreal, comprising 48 iRBD patients, 29 PD patients, and 41 healthy controls. Within the iRBD cohort, 16 patients were diagnosed with Mild Cognitive Impairment (MCI), providing a basis for subtyping analysis.

To capture a comprehensive picture of brain morphology, the researchers extracted four distinct metrics:

  1. Cortical Thickness: The depth of the gray matter.
  2. Surface Area: The expansion or contraction of cortical folds.
  3. Volume: The total space occupied by specific regions.
  4. Deformation-Based Morphometry (DBM): A technique quantifying the local "stretching" or "shrinking" required to map an individual brain to a standard template.

The machine learning pipeline involved a Random Forest classifier for feature selection from 287 brain regions, followed by training across six different architectures, including Support Vector Machines (SVM) and K-Nearest Neighbors (KNN). The models were validated using 25 unique, stratified train-test splits to ensure generalizability.

Results and Classification Performance

The ML models demonstrated robust discriminative power across three primary tasks:

  • iRBD vs. Controls: Achieved 79.6% accuracy. The most predictive features were the deformation of the caudal middle frontal gyrus and putamen, thinning of the superior frontal gyrus, and reduced volume of the inferior parietal cortex and insula.
  • iRBD vs. Parkinson’s Disease: Achieved 82.0% accuracy. The model successfully distinguished these groups based on smaller insular volume and surface area in iRBD patients, alongside specific thinning patterns in the entorhinal cortex.
  • iRBD with MCI vs. without MCI: Achieved 84.8% accuracy. This subtyping task relied on the thinning of the pars triangularis and superior temporal gyrus, as well as deformation of the isthmus of the cingulate gyrus.

Discussion: The Power of DBM and Regional Vulnerability

A significant finding of the study was that Deformation-Based Morphometry (DBM) outperformed traditional cortical thickness as a predictor. DBM is particularly sensitive to changes in subcortical structures like the putamen and thalamus, as well as subtle white matter contractions that surface-based metrics often miss.

The insula emerged as a critical differentiator, particularly for distinguishing iRBD from PD. Because insular atrophy is closely linked to cognitive decline and DLB, this suggests that MRI-based ML could help identify which iRBD patients are on a "dementia-first" trajectory. Additionally, the consistent identification of the caudal middle frontal gyrus across tasks reinforces the theory that frontal executive centers are among the first to undergo structural changes in synucleinopathies.

Conclusion

This research demonstrates that morphology-based machine learning can effectively identify and subtype iRBD, potentially serving as a non-invasive screening tool for neurodegenerative diseases. While the results are promising, future work requires larger, multicentric datasets to ensure these models generalize across different populations and scanner types. Longitudinal data will also be essential to confirm if these structural signatures can accurately predict the specific timing of conversion to Parkinson's or dementia. Combining these MRI signatures with other biomarkers remains the ultimate goal for early clinical intervention.

Summary updated

Abstract

Abstract Background Idiopathic rapid eye movement sleep behavior disorder (iRBD) is a major risk factor for synucleinopathies, and patients often present with clinical signs and morphological brain changes. However, there is a heterogeneity in the presentation and progression of these alterations, and brain regions that are more vulnerable to neurodegeneration remain to be determined. Objectives To assess the feasibility of morphology-based machine learning in the identification and subtyping of iRBD. Methods For the classification tasks [iRBD (n=48) vs controls (n=41); iRBD vs Parkinson’s disease (n=29); iRBD with mild cognitive impairment (n=16) vs without mild cognitive impairment (n=32)], machine learning models were trained with morphometric measurements (thickness, surface area, volume, and deformation) extracted from T1-weighted structural magnetic resonance imaging. Model performance and the most discriminative brain regions were analyzed and identified. Results A high accuracy was reported for iRBD vs controls (79.6%, deformation of the caudal middle frontal gyrus and putamen, thinning of the superior frontal gyrus, and reduced volume of the inferior parietal cortex and insula), iRBD vs Parkinson’s disease (82%, smaller volume and surface area of the insula, lower thinning of the entorhinal cortex and lingual gyrus, and greater volume of the fusiform gyrus), and iRBD with vs without mild cognitive impairment (84.8%, thinning of the pars triangularis, superior temporal gyrus, transverse temporal cortex, larger surface area of the superior temporal gyrus, and deformation of isthmus of the cingulate gyrus). Conclusions Morphology-based machine learning approaches may allow for detection and subtyping of iRBD, potentially enabling efficient preclinical identification of synucleinopathies.
BibTeX
@misc{mei2021identification,
  author = {Jie Mei and Shady Rahayel and Christian Desrosiers and Ronald B. Postuma and Jacques Montplaisir and Julie Carrier and Oury Monchi and Johannes Frasnelli and Jean‐François Gagnon},
  title = {{Identification of REM Sleep Behavior Disorder by Magnetic Resonance Imaging and Machine Learning}},
  year = {2021},
  doi = {10.1101/2021.09.18.21263779},
  howpublished = {medRxiv},
}
Figure