Three-fold cross-validation of parkinsonian brain patterns

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Neurology

Background:

• Abnormal brain physiological networks are implicated in various neurological diseases.
• Multivariate computational algorithms, particularly principal component analysis (PCA), can identify these networks from functional imaging data.

Purpose of the Study:

• To demonstrate the reproducibility of brain network patterns derived from positron emission tomography (PET) data in independent cohorts of parkinsonian patients.
• To validate these network patterns for differential diagnosis of idiopathic Parkinson's disease (iPD), multiple system atrophy (MSA), and progressive supranuclear palsy (PSP).

Main Methods:

• Applied principal component analysis (PCA) to functional PET imaging data from iPD, MSA, and PSP patients.
• Correlated voxel values of network patterns across independent datasets to assess reproducibility.
• Utilized a logistic regression classification algorithm with three-fold cross-validation for differential diagnosis.

Main Results:

• High correlation of voxel values for network patterns derived from the same condition in different datasets.
• Achieved high accuracy (82%-93%) in training sets and approximately 81% accuracy for prospective test subjects in differential diagnosis.
• Demonstrated the utility of network scores for single-subject differential diagnosis.

Conclusions:

• Brain network patterns identified via PET and PCA are reproducible across independent patient populations.
• These reproducible network patterns serve as reliable biomarkers for the differential diagnosis of distinct parkinsonian syndromes.
• The developed classification algorithm shows significant potential for clinical application in diagnosing parkinsonian disorders.