Healthy core: Harmonizing brain MRI for supporting multicenter migraine classification studies.

Area of Science:

• Neuroimaging
• Machine Learning
• Data Science

Background:

• Multicenter neuroimaging studies are crucial for large sample sizes but face challenges with data heterogeneity.
• Varying scanners and protocols across sites can hinder the generalizability of machine learning models.
• Ensuring reproducible results requires robust classification models applicable to diverse datasets.

Purpose of the Study:

• To enhance the generalizability of machine learning models for classifying migraine patients and healthy controls using brain MRI data.
• To propose and validate a data harmonization strategy by identifying a 'healthy core' of homogeneous controls.
• To improve the performance of predictive models on unseen data from different scanners or centers.

Main Methods:

• Utilized Maximum Mean Discrepancy (MMD) within Geodesic Flow Kernel (GFK) space to quantify dataset variability.
• Implemented a 'healthy core' strategy by selecting homogeneous healthy control subjects from multicenter datasets.
• Developed and evaluated classification models using harmonized and non-harmonized datasets.

Main Results:

• The 'healthy core' approach effectively mitigated data heterogeneity from multicenter, multi-scanner studies.
• A homogeneous dataset derived from healthy controls significantly improved classification accuracy.
• Accuracy for classifying both episodic and chronic migraineurs saw a notable 25% improvement.

Conclusions:

• Leveraging a 'healthy core' is a beneficial strategy for improving the generalizability of neuroimaging-based machine learning models.
• Data harmonization through identifying homogeneous control groups enhances model performance and reproducibility.
• This method holds promise for developing more reliable diagnostic and predictive tools for neurological conditions like migraine.