Classification of first-episode psychosis using cortical thickness: A large multicenter MRI study

Area of Science:

• Neuroimaging
• Machine Learning
• Psychiatry

Background:

• Machine learning in first-episode psychosis (FEP) primarily used brain volumes.
• Previous cortical thickness studies often used single-site data, limiting generalizability.
• Parcellation approaches may oversimplify complex cortical patterns.

Purpose of the Study:

• To conduct a large-scale, multi-site analysis of cortical thickness for FEP classification.
• To compare parcellation versus vertex-wise approaches for neuroimaging classification.
• To investigate the impact of demographic and site-specific variables on classification accuracy and generalizability.

Main Methods:

• Utilized structural MRI data from 428 FEP subjects and 448 healthy controls across 8 centers.
• Extracted cortical thickness using both parcellation (68 areas) and vertex-wise (20,484 vertices) methods.
• Employed Linear Support Vector Machine within a nested cross-validation framework, stratifying by MRI scanner for generalizability.

Main Results:

• Vertex-wise cortical thickness maps achieved higher classification accuracy (66.2% balanced accuracy, 72% AUC) than parcellation methods.
• Stratification by MRI scanner improved across-site generalizability.
• Temporal brain regions were most influential for classification, and predictive scores correlated with clinical variables like age at onset and symptom severity.

Conclusions:

• Temporal cortical thickness can classify individuals with FEP from healthy controls, though clinical relevance requires further investigation.
• Multi-site analysis and accounting for site-dependent variables enhance the generalizability of machine learning models in neuroimaging.
• Vertex-wise analysis offers a more detailed and accurate approach compared to traditional parcellation methods for FEP classification.