Machine Learning-Driven Approach to Identify Freezing of Gait in Individuals with Parkinson's Disease Using Conventional Structural MRI and Clinical Measures.
Authors
Affiliations (2)
Affiliations (2)
- From the Department of Radiology (G.N.R., A.J.G., V.R.M.), University of Alabama at Birmingham, Birmingham, Alabama, United States; Department of Physical Therapy and Athletic Training (J.K.L.), Saint Louis University, St. Louis, Missouri, United States; and Lou Ruvo Center for Brain Health (Z.M.), Cleveland Clinic Foundation, Las Vegas, Nevada, United States.
- From the Department of Radiology (G.N.R., A.J.G., V.R.M.), University of Alabama at Birmingham, Birmingham, Alabama, United States; Department of Physical Therapy and Athletic Training (J.K.L.), Saint Louis University, St. Louis, Missouri, United States; and Lou Ruvo Center for Brain Health (Z.M.), Cleveland Clinic Foundation, Las Vegas, Nevada, United States. [email protected].
Abstract
Freezing of gait (FOG) presents a significant challenge in the management of Parkinson's disease (PD). Our study explored the potential to predict PD-FOG using an unbiased machine learning (ML) approach that leverages conventional T1-weighted MRI and clinical measures. Thirty-seven participants (16 PD-FOG, 21 PD-nFOG) underwent standard isotropic 1mm³ T1-weighted MRI. Brain morphometric measures, including subcortical volume, cortical volume, mean curvature, area, local gyrification index, and thickness, were extracted using FreeSurfer7. Participants were divided into discovery (13 PD-FOG, 17 PD-nFOG) and independent testing (3 PD-FOG, 4 PD-nFOG) datasets. We tested the predictive ability of each FreeSurfer-derived measure, each clinical measure, and every combination of those measures using three ML models: Random Forest (RF), Support Vector Machine (SVM)-Linear, and SVM-Non-Linear. Feature reduction was performed using the least absolute shrinkage and selection operator before model development. The SVM-linear model outperformed SVM-Non-Linear and RF models when tested on the independent dataset (area under the curve [AUC]: 0.71, precision: 75%, sensitivity: 75%, specificity: 66.67%). FreeSurfer-derived cortical area from twenty-seven regions predicted PD-FOG, involving several cortical and subcortical regions. None of the other measures, either in combination or isolation, predicted PD-FOG. The identified features were significantly correlated with clinical and physical therapy measures of PD-FOG using univariate and multivariate statistics, bolstering confidence in the selected feature set. Our results demonstrate that FreeSurfer-derived cortical area measures from 27 key regions across the frontal, temporal, parietal, and occipital lobes can moderately predict FOG in PD (AUC = 0.71) using a linear SVM model. While preliminary, our work outlines an MRI-based analytical approach that may inform future external validation efforts and contribute to understanding the potential role of cortical morphology in PD-FOG risk. However, given the limited sample size and constrained independent testing cohort, these findings should be interpreted as exploratory and warrant replication in larger, multi-center studies.