An uncertainty-aware, shareable, and transparent neural network architecture for brain-age modeling

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomarkers

Background:

• Brain age prediction from neuroimaging is a key biomarker for brain changes across disorders.
• Current machine learning models lack uncertainty quantification, leading to biased results.
• Existing models often use homogeneous, non-validated datasets and face data sharing restrictions.

Purpose of the Study:

• To develop an uncertainty-aware, shareable, and transparent AI model for biological age estimation.
• To address limitations of existing models in handling data variability and ensuring reproducibility.
• To improve the detection of deviant brain aging patterns.

Main Methods:

• Developed a Monte Carlo dropout composite quantile regression (MCCQR) Neural Network.
• Trained the MCCQR model on a large dataset (N=10,691) from the German National Cohort.
• Implemented robust, distribution-free uncertainty quantification for high-dimensional neuroimaging data.

Main Results:

• The MCCQR model achieved lower error rates compared to existing methods.
• Demonstrated prevention of spurious associations in brain aging analysis.
• Showcased increased statistical power for detecting deviant brain aging.

Conclusions:

• The MCCQR model offers a more reliable approach to biological age estimation.
• The model's uncertainty quantification enhances the robustness of neuroimaging findings.
• Publicly releasing the model and code promotes transparency and further research in brain aging.