Region-wise stacking ensembles for estimating brain-age using structural MRI.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Predictive modeling of brain aging using structural MRI is crucial for understanding healthy and disease-related aging.
• High-dimensional MRI data present challenges in model generalizability, interpretability, and data privacy.
• Current methods like voxel averaging reduce anatomical specificity and can lead to information loss.

Purpose of the Study:

• To develop and evaluate a novel two-level stacking ensemble (SE) approach for brain-aging prediction.
• To improve prediction accuracy and data privacy in structural MRI analysis.
• To explore the biological insights and clinical utility of the proposed SE model.

Main Methods:

• A two-level stacking ensemble (SE) model was developed, with regional models predicting age from voxel-wise data in the first level.
• A second-level model fused these regional predictions for final age estimation.
• Eight data fusion scenarios were tested using Gray Matter Volume (GMV) from four large datasets.

Main Results:

• The SE approach significantly outperformed baseline region-wise averaging methods in predicting age (MAE = 4.75 vs 5.68).
• Optimal performance was achieved using out-of-sample predictions for regional models and site-specific training for the second-level model.
• The SE model demonstrated improved robustness, enhanced data privacy, and provided new biological insights into the aging process.

Conclusions:

• The stacking ensemble (SE) model offers a superior alternative to traditional methods for brain-aging prediction using structural MRI.
• The SE approach enhances predictive accuracy while preserving or improving data privacy.
• The model shows promise for clinical applications, including the identification of accelerated aging in neurodegenerative diseases like Alzheimer's.