Mitigating site effects in covariance for machine learning in neuroimaging data.

Area of Science:

• Neuroscience
• Medical Imaging
• Data Science

Background:

• Multi-site neuroimaging studies are crucial for large-scale neuroscience research but suffer from site-specific data variations.
• These variations can bias comparisons, obscure true biological signals, and introduce false findings.
• Existing harmonization methods focus on mean and variance, potentially neglecting other critical data differences.

Purpose of the Study:

• To evaluate the effectiveness of current data harmonization techniques for machine learning (ML) in neuroimaging.
• To introduce a novel harmonization method that addresses site-specific differences in covariance.
• To improve the reliability and accuracy of ML models in multi-site neuroimaging studies.

Main Methods:

• Analysis of data from the Alzheimer's Disease Neuroimaging Initiative to identify cross-site covariance differences.
• Development and application of Correcting Covariance Batch Effects (CovBat), a novel harmonization technique.
• Comparison of CovBat with existing methods in terms of harmonizing correlation matrices and impact on ML model performance.

Main Results:

• Popular harmonization methods fail to account for site-specific covariance variations.
• CovBat successfully harmonizes within-site correlation matrices, addressing mean, variance, and covariance.
• ML models trained on CovBat-harmonized data show reduced bias (e.g., inability to distinguish scanner manufacturer) while maintaining accurate disease group prediction.

Conclusions:

• Harmonizing covariance is essential for effective machine learning in multi-site neuroimaging.
• CovBat offers a robust solution for addressing complex site effects beyond mean and variance.
• The proposed method enhances the utility of multi-site neuroimaging data for advancing neuroscience research.