ComBat-Predict Enhances Generalizability of Neuroimaging Models to New Sites.

Area of Science:

• Neuroimaging
• Biostatistics
• Computational Neuroscience

Background:

• Multi-site neuroimaging studies are crucial for understanding brain development and neurodegeneration.
• Site-specific variations (site effects) in imaging data can bias results.
• Existing harmonization methods often fail to generalize to new, unseen sites.

Purpose of the Study:

• To develop a novel harmonization method, ComBat-Predict (CB-Predict), that generalizes to new sites.
• To address limitations of current methods in handling smaller sample sizes and unknown site effects at new locations.
• To improve the translation of neuroimaging models to diverse datasets and clinical practice.

Main Methods:

• Proposed ComBat-Predict (CB-Predict) method, an extension of the ComBat algorithm.
• Applied CB-Predict to data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) for cortical thickness prediction.
• Validated CB-Predict's ability to reduce site-related variance using data from the Lifespan Brain Chart Consortium.

Main Results:

• CB-Predict demonstrated high accuracy in predicting cortical thickness measures when generalizing to new data.
• The method effectively mitigated site-related bias in neuroimaging data.
• CB-Predict successfully reduced variance in centile scores derived from multi-site data.

Conclusions:

• ComBat-Predict (CB-Predict) offers a robust solution for harmonizing neuroimaging data from new sites.
• The method enhances the generalizability of neuroimaging models, facilitating wider research and clinical applications.
• CB-Predict enables effective translation of findings across different study sites and populations.