Harmonization for Parkinson's Disease Multi-Dataset T1 MRI Morphometry Classification.

Area of Science:

• Neuroimaging
• Machine Learning
• Neurodegenerative Diseases

Background:

• Personalized disease classification using structural MRI morphometry offers clinical advantages over traditional statistics.
• Training classifiers on multi-scanner neuroimaging datasets is challenged by confounding batch effects.

Purpose of the Study:

• To evaluate the ComBat harmonization model for classifying Parkinson's disease (PD) from healthy controls using multi-scanner MRI data.
• To identify common pitfalls, such as data leakage, in multi-scanner neuroimaging classification pipelines.

Main Methods:

• Utilized a cohort of 372 subjects (216 PD, 156 controls) across 11 MRI scanners.
• Extracted FreeSurfer and Jacobian determinant morphometry data.
• Compared single-scanner and multi-scanner classification pipelines, assessing the impact of batch effect harmonization and data leakage prevention.

Main Results:

• Single-scanner classifiers showed highly variable performance (mean AUC: 0.651 ± 0.144).
• Multi-scanner classifiers incorporating batch effects achieved high AUC (0.902), but pipelines preventing data leakage performed poorly (AUC: 0.550).
• Batch effects significantly impact classification generalizability across scanners.

Conclusions:

• Batch effects are a critical challenge for neuroimaging-based disease classification, limiting the generalizability of single-scanner models.
• Effective harmonization strategies must avoid circularity to prevent reporting overly optimistic results in classifier development.