Extensive Evaluation of Morphological Statistical Harmonization for Brain Age Prediction.

Area of Science:

• Neuroimaging
• Brain Development
• Computational Neuroscience

Background:

• Understanding brain structural changes across the lifespan is vital for identifying normal development and atypical patterns.
• Open access initiatives have advanced brain structure characterization by enabling multi-site data sharing.
• Harmonizing multi-site neuroimaging data is essential to mitigate bias and enhance statistical power.

Purpose of the Study:

• To evaluate three data harmonization techniques for age prediction using the Autism Brain Imaging Data Exchange (ABIDE) dataset.
• To assess the impact of harmonization on machine learning model performance and feature relevance for age prediction in controls and autism spectrum disorder (ASD) subjects.
• To propose a stability index for robust feature selection and clinical validation across harmonization strategies.

Main Methods:

• Extracted morphological features from T1-weighted MRI scans of 654 subjects across 17 sites.
• Employed three machine learning regression models for biological age prediction.
• Developed a framework to quantify harmonization effects on model performance and age-predictive brain regions.

Main Results:

• Two harmonization strategies showed comparable predictive model accuracy.
• Significant discrepancies were observed in age-related predictive brain regions between harmonization methods, particularly for ASD subjects.
• The proposed stability index aids in identifying consistent and meaningful features.

Conclusions:

• Harmonization strategy choice impacts the identification of neurobiological correlates of aging, especially in clinical populations like ASD.
• A stability-based approach is recommended for robust feature selection and clinical validation in multi-site neuroimaging studies.
• This work provides a framework for optimizing data harmonization in large-scale brain imaging research.