Harmonizing and aligning M/EEG datasets with covariance-based techniques to enhance predictive regression modeling.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Neuroscience studies require large datasets for machine learning (ML), but data heterogeneity (dataset shifts) across sites, devices, or protocols limits ML model generalization.
• Magnetoencephalography (M/EEG) data are often represented by covariance matrices, which are sensitive to dataset shifts caused by variations in brain activity, anatomy, or equipment.

Purpose of the Study:

• To investigate the impact of dataset shifts on M/EEG data and evaluate domain adaptation methods for improving ML model generalization.
• To theoretically explain how anatomical, physiological, and technical factors induce covariance shifts in M/EEG data.

Main Methods:

• Developed and evaluated model-based dataset alignment methods (re-centering, re-scaling, rotation correction) leveraging the geometry of covariance matrices.
• Applied and assessed these alignment methods on MEG and EEG datasets for brain age prediction tasks.
• Utilized controlled simulations to theoretically validate the alignment methods' effectiveness in addressing covariance shifts.

Main Results:

• Paired rotation correction was crucial for M/EEG data from the same subjects under different tasks (e.g., rest vs. passive/smt tasks), improving R-squared by up to 0.17.
• Re-centering improved performance when including unseen subjects and different tasks (e.g., +0.096 R-squared for rest-passive).
• Re-centering was essential for generalizing brain age prediction to independent datasets from different populations and devices, achieving performance close to within-dataset predictions.

Conclusions:

• Domain adaptation techniques, particularly re-centering and rotation correction, can significantly enhance the generalization of M/EEG-based regression models across diverse datasets.
• Statistically harmonizing M/EEG data through alignment methods is critical for successful cross-dataset ML applications in neuroscience.
• This approach facilitates the integration of large public datasets, overcoming limitations imposed by dataset shifts.