Towards a Foundation Model for Brain Age Prediction using coVariance Neural Networks

TL;DR

This work introduces NeuroVNN, a coVariance Neural Network-based foundation model for brain age prediction that operates on covariance matrices derived from cortical thickness features. Trained on a healthy adult cohort, NeuroVNN demonstrates scale-free transferability across brain atlases and remains anatomically interpretable through a non-adaptive readout, enabling region-wise contributions to the age estimate. Fine-tuning with age-bias correction yields biologically meaningful $\Delta$-Age metrics, which correlate with dementia-related biomarkers such as Alzheimer's progression score ($\text{APS}$) and, in amyloid-positive individuals, plasma NfL, thereby supporting applicability to presymptomatic and neurodegenerative contexts. The results suggest a robust, atlas-agnostic foundation model for brain aging analysis with potential for biomarker discovery and cross-context applicability in computational neuroscience.

Abstract

Brain age is the estimate of biological age derived from neuroimaging datasets using machine learning algorithms. Increasing brain age with respect to chronological age can reflect increased vulnerability to neurodegeneration and cognitive decline. In this paper, we study NeuroVNN, based on coVariance neural networks, as a paradigm for foundation model for the brain age prediction application. NeuroVNN is pre-trained as a regression model on healthy population to predict chronological age using cortical thickness features and fine-tuned to estimate brain age in different neurological contexts. Importantly, NeuroVNN adds anatomical interpretability to brain age and has a `scale-free' characteristic that allows its transference to datasets curated according to any arbitrary brain atlas. Our results demonstrate that NeuroVNN can extract biologically plausible brain age estimates in different populations, as well as transfer successfully to datasets of dimensionalities distinct from that for the dataset used to train NeuroVNN.

Figures (12)

• Figure 1: An overview of NeuroVNN as the paradigm for foundation model for brain age prediction.
• Figure 2: Age distribution of dataset used to train NeuroVNN.
• Figure 3: One-shot transferability of NeuroVNN for chronological age prediction task. NeuroVNN was trained on dataset curated according to 100 parcellations version of Schaefer's brain atlas and transferred to other versions of the dataset curated according to different brain atlases.
• Figure 4: Age distributions in PREVENT-AD dataset and HC group of ADNI-1 dataset.
• Figure 5: Results on PREVENT-AD dataset obtained by NeuroVNN (trained on dataset with $100$ cortical thickness features). The figures in the right column plot the magnitude of positive correlation between APS and the associated entry at the final layer of NeuroVNN on a brain region, thus, providing an anatomic signature to the findings in the left column.