Abstract:Graph convolutional networks (GCN) leverage topology-driven graph convolutional operations to combine information across the graph for inference tasks. In our recent work, we have studied GCNs with covariance matrices as graphs in the form of coVariance neural networks (VNNs) that draw similarities with traditional PCA-driven data analysis approaches while offering significant advantages over them. In this paper, we first focus on theoretically characterizing the transferability of VNNs. The notion of transferability is motivated from the intuitive expectation that learning models could generalize to "compatible" datasets (possibly of different dimensionalities) with minimal effort. VNNs inherit the scale-free data processing architecture from GCNs and here, we show that VNNs exhibit transferability of performance over datasets whose covariance matrices converge to a limit object. Multi-scale neuroimaging datasets enable the study of the brain at multiple scales and hence, can validate the theoretical results on the transferability of VNNs. To gauge the advantages offered by VNNs in neuroimaging data analysis, we focus on the task of "brain age" prediction using cortical thickness features. In clinical neuroscience, there has been an increased interest in machine learning algorithms which provide estimates of "brain age" that deviate from chronological age. We leverage the architecture of VNNs to extend beyond the coarse metric of brain age gap in Alzheimer's disease (AD) and make two important observations: (i) VNNs can assign anatomical interpretability to elevated brain age gap in AD, and (ii) the interpretability offered by VNNs is contingent on their ability to exploit specific principal components of the anatomical covariance matrix. We further leverage the transferability of VNNs to cross validate the above observations across different datasets.

What problem does this paper attempt to address?

The main problems that this paper attempts to solve include: 1. **Theoretical analysis of transferability**: - The author studied the transferability of coVariance Neural Networks (VNNs) based on covariance matrices between datasets of different dimensions. Specifically, they attempted to prove that VNNs can be transferred from a dataset with dimension \(m_1\) to another dataset with dimension \(m_2\) without retraining and maintain the same or close performance. - This property is particularly useful for handling multi - scale data, for example, in neuroimaging, different brain atlases can provide different - resolution representations of the same information. 2. **Interpretable brain - age prediction based on anatomical features**: - The paper explored the task of brain - age prediction using VNNs, especially predicting an individual's "brain - age" (i.e., biological age) through cortical thickness features. The difference between brain - age and chronological age (called the brain - age gap, \(\Delta\)-Age) can reflect an individual's susceptibility to neurodegenerative diseases or cognitive impairments. - The author emphasized that existing machine - learning methods often lack the ability to explain the contributions of specific brain regions when predicting brain - age. VNNs can identify brain regions that lead to an increased brain - age gap by analyzing specific principal components, thus providing an anatomical explanation. ### Specific problems and solutions - **Transferability problem**: - **Challenge**: Traditional statistical models such as PCA cannot be directly applied to datasets of different dimensions because principal components depend on the feature space of the original dataset and need to be re - evaluated when the number of features changes. - **Solution**: VNNs operate on covariance matrices instead of directly on feature vectors, enabling the model to be transferred between datasets of different dimensions without retraining. This provides theoretical support for handling multi - scale data. - **Explanatory problem in brain - age prediction**: - **Challenge**: Existing methods usually only focus on prediction accuracy when predicting brain - age and ignore the explanation of the contributions of specific brain regions. - **Solution**: VNNs can identify brain regions that contribute to an increased brain - age gap by analyzing specific eigenvectors of the covariance matrix, thus providing an anatomical explanation. In addition, VNNs can also verify the robustness of these findings in datasets of different scales. ### Theoretical contributions - **Theoretical framework**: The author proposed the theoretical conditions for the transferability of VNNs between datasets of different dimensions. Specifically, when two covariance matrices approach the same limit object under certain conditions, the outputs of VNNs are close in a certain sense. - **Experimental verification**: Experiments on multi - scale neuroimaging datasets proved the transferability and anatomical explanation ability of VNNs. In particular, in Alzheimer's disease (AD) patients, VNNs can identify specific brain regions that lead to an increased brain - age gap. In conclusion, this paper aims to solve the deficiencies of existing methods in transferability and interpretability by theoretically analyzing and experimentally proving the advantages of VNNs in multi - scale data processing, especially in the application of brain - age prediction tasks.

Transferability of coVariance Neural Networks and Application to Interpretable Brain Age Prediction using Anatomical Features

Predicting Brain Age using Transferable coVariance Neural Networks

Explainable Brain Age Prediction using coVariance Neural Networks

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

Regional Brain Fusion: Graph Convolutional Network for Alzheimer's Disease Prediction and Analysis

Transferability of Brain decoding using Graph Convolutional Networks

Sparse Covariance Neural Networks

Interpretable Graph Convolutional Network Of Multi-Modality Brain Imaging For Alzheimer’s Disease Diagnosis

Multi-Modal Diagnosis of Alzheimer's Disease using Interpretable Graph Convolutional Networks

A voxel-level approach to brain age prediction: A method to assess regional brain aging

Transfer Learning Models on Brain Age Prediction

Whole-brain Prediction Analysis with GraphNet

Fair CoVariance Neural Networks

Ensemble Manifold Regularized Multi-Modal Graph Convolutional Network for Cognitive Ability Prediction

Early prediction of dementia using fMRI data with a graph convolutional network approach

Spatiotemporal Covariance Neural Networks

S-GCN: A Siamese Spectral Graph Convolutions on Brain Connectivity Networks

Beyond the Snapshot: Brain Tokenized Graph Transformer for Longitudinal Brain Functional Connectome Embedding

MC-RVAE: Multi-channel recurrent variational autoencoder for multimodal Alzheimer's disease progression modelling

A Graph Convolutional Network Based on Univariate Neurodegeneration Biomarker for Alzheimer's Disease Diagnosis

DS-GCNs: Connectome Classification using Dynamic Spectral Graph Convolution Networks with Assistant Task Training