Spectral Graph Sample Weighting for Interpretable Sub-cohort Analysis in Predictive Models for Neuroimaging.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Informatics

Background:

• Brain disorders exhibit significant heterogeneity in mechanisms, development, and severity.
• This heterogeneity, influenced by factors like sex and genetics, impacts the predictive accuracy of machine learning models.
• Existing methods struggle to effectively model and address this patient variability.

Purpose of the Study:

• To develop a novel sample weighting scheme for machine learning models to address heterogeneity in brain disorder prediction.
• To improve the interpretability and predictive power of models by accounting for subject-specific factors.
• To identify sub-cohorts with varying degrees of predictability within patient populations.

Main Methods:

• Proposed a method to model subject weights as a linear combination of spectral population graph eigenbases.
• Captured similarity of demographic and disease-related factors across subjects using a graph.
• Applied the weighting scheme to predict heavy alcohol drinking initiation and differentiate Dementia from Mild Cognitive Impairment.

Main Results:

• The proposed sample weighting scheme improved interpretability compared to existing methods.
• Successfully highlighted sub-cohorts with distinct characteristics and varying model accuracy.
• Demonstrated effectiveness in predicting alcohol drinking initiation and detecting cognitive impairment.

Conclusions:

• The developed sample weighting strategy effectively models patient heterogeneity in machine learning for brain disorders.
• This approach enhances model interpretability and identifies specific patient subgroups with differential predictability.
• The method shows promise for improving diagnostic and prognostic accuracy in neurological and psychiatric conditions.