A comparative machine learning study of schizophrenia biomarkers derived from functional connectivity

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Functional connectivity is a promising biomarker for schizophrenia.
• High dimensionality and small sample sizes risk overfitting predictive models.
• Low-dimensional connectome representations like cortical gradients are proposed but their predictive power for schizophrenia is unclear.

Purpose of the Study:

• To evaluate which connectome features—functional connectivity, gradients, or gradient dispersion—best identify schizophrenia.
• To compare the predictive capacity of these different feature types.

Main Methods:

• Utilized resting-state functional MRI data from 936 individuals across three datasets (COBRE, LA5c, SRPBS-1600).
• Developed a computational pipeline to assess over a million features.
• Selected top 1% of features by permutation importance and trained 13 classifiers using 10-fold cross-validation.

Main Results:

• Raw functional connectivity significantly outperformed low-dimensional derivatives (cortical gradients and gradient dispersion) in identifying schizophrenia.
• Edges connecting primary sensory regions demonstrated the highest classification performance.
• Functional connectivity within primary sensory regions showed superior discrimination between schizophrenia patients and controls.

Conclusions:

• Raw functional connectivity is a more effective biomarker for schizophrenia than its low-dimensional derivatives.
• Primary sensory region connectivity is crucial for discriminating schizophrenia.
• The proposed feature selection pipeline can aid future research on schizophrenia subtypes and transdiagnostic phenomena.