Classification of first-episode psychosis: a multi-modal multi-feature approach integrating structural and diffusion...

Area of Science:

• Neuroimaging
• Machine Learning
• Psychiatry

Background:

• Most psychosis classification studies focus on chronic patients using single machine learning methods.
• First-episode psychosis (FEP) classification remains challenging, necessitating advanced analytical approaches.

Purpose of the Study:

• To compare different machine learning classification methods for FEP using multi-modal imaging data.
• To identify reliable neuroimaging biomarkers for early psychosis detection.

Main Methods:

• Utilized structural MRI and diffusion tensor imaging in 23 FEP patients and 23 healthy controls.
• Implemented a multivariate multiple kernel learning (MKL) approach, comparing it with support vector machines.
• Analyzed cortical and subcortical structures and white matter fiber bundles.

Main Results:

• Achieved over 90% classification accuracy between FEP patients and healthy individuals.
• Multiple kernel learning outperformed support vector machines in classification performance.
• Identified specific regions (middle/superior frontal gyrus, parahippocampal gyrus) and white matter tracts (uncinate fascicles, cingulum) with >70% accuracy.

Conclusions:

• Multivariate machine learning integrating multi-modal imaging data reliably classifies FEP.
• Specific grey and white matter structures show high classification reliability, suggesting an impaired prefronto-limbic network in FEP, particularly in the right hemisphere.