Data Driven Classification Using fMRI Network Measures: Application to Schizophrenia.

Area of Science:

• Neuroimaging
• Biomarker Discovery
• Machine Learning in Neuroscience

Background:

• Functional magnetic resonance imaging (fMRI) is used to identify brain disorder biomarkers.
• Classification pipelines typically involve feature extraction from time series for subject classification.
• Identifying robust biomarkers is crucial for understanding and diagnosing brain disorders.

Purpose of the Study:

• To compare single and double cross-validation schemes for biomarker identification in schizophrenia.
• To evaluate the impact of functional vs. anatomical atlases on classification performance.
• To assess the effect of prewhitening on feature extraction and classification accuracy.

Main Methods:

• Utilized graph theoretic measures as features from fMRI time series.
• Compared classification performance using single and double cross-validation on 170 subjects (schizophrenia patients and healthy controls).
• Investigated the influence of functional and anatomical brain atlases and prewhitening.

Main Results:

• Double cross-validation decreased classification performance by 20% compared to single cross-validation.
• Anatomical atlases yielded higher classification results than functional atlases.
• Prewhitening improved classification performance by 10%.

Conclusions:

• A classification performance of 80% was achieved using double cross-validation, prewhitened time series, and an anatomical atlas.
• Low within-subject reproducibility across scans suggests significant influence of subject state on fMRI features and classification.
• Findings highlight the importance of cross-validation strategy and atlas choice in biomarker discovery, while cautioning on the impact of transient subject states.