Stability Evaluation of Brain Changes in Parkinson's Disease Based on Machine Learning

Area of Science:

• Neuroimaging
• Machine Learning
• Biomarker Discovery

Background:

• Structural MRI (sMRI) is used to study cerebral changes in Parkinson's disease (PD).
• Previous studies focused on group-level changes and yielded inconsistent results, hindering identification of reliable PD biomarkers.
• A need exists for individual-level analysis and robust identification of brain regions associated with PD.

Purpose of the Study:

• To develop a machine learning approach for individual-level PD diagnosis using sMRI data.
• To identify minimal, non-redundant, and informative brain features (from gray and white matter) as potential PD biomarkers.
• To validate the stability and consistency of identified brain regions with existing literature.

Main Methods:

• Employed four feature selection methods (ReliefF, graph-theory, RFE, stability selection) on gray matter (GM) and white matter (WM) data.
• Utilized a support vector machine (SVM) to build classification models for distinguishing PD patients from healthy controls (HCs).
• Conducted horizontal (cross-method) and vertical (comparison with conventional analysis) analyses to assess biomarker stability.

Main Results:

• Achieved high classification performance: up to 89.58% accuracy for GM and 71.18% for WM.
• Identified a minimal set of informative brain regions, with most aligning with previous PD research findings.
• Confirmed specific brain abnormalities in the superior frontal gyrus (SFGdor) and lingual gyrus (LING) as potential PD biomarkers.

Conclusions:

• Machine learning techniques effectively enable individual-level PD prediction and biomarker identification from sMRI data.
• The identified brain regions, particularly SFGdor and LING, are robust potential biomarkers for Parkinson's disease.
• This approach offers a valuable decision support system for clinicians in diagnosing Parkinson's disease.