Spectral Graph Convolutions for Population-based Disease Prediction

Abstract: Exploiting the wealth of imaging and non-imaging information for disease prediction tasks requires models capable of representing, at the same time, individual features as well as data associations between subjects from potentially large populations. Graphs provide a natural framework for such tasks, yet previous graph-based approaches focus on pairwise similarities without modelling the subjects' individual characteristics and features. On the other hand, relying solely on subject-specific imaging feature vectors fails to model the interaction and similarity between subjects, which can reduce performance. In this paper, we introduce the novel concept of Graph Convolutional Networks (GCN) for brain analysis in populations, combining imaging and non-imaging data. We represent populations as a sparse graph where its vertices are associated with image-based feature vectors and the edges encode phenotypic information. This structure was used to train a GCN model on partially labelled graphs, aiming to infer the classes of unlabelled nodes from the node features and pairwise associations between subjects. We demonstrate the potential of the method on the challenging ADNI and ABIDE databases, as a proof of concept of the benefit from integrating contextual information in classification tasks. This has a clear impact on the quality of the predictions, leading to 69.5% accuracy for ABIDE (outperforming the current state of the art of 66.8%) and 77% for ADNI for prediction of MCI conversion, significantly outperforming standard linear classifiers where only individual features are considered.

• The paper proposes an innovative GCN framework that fuses imaging features with phenotypic data in sparse graphs for improved disease prediction.
• It employs spectral graph convolutions via Chebyshev polynomials to capture both local and global contextual relationships among subjects.
• The method demonstrates significant gains with 69.5% accuracy on ABIDE and 77% on ADNI, surpassing previous state-of-the-art results.

Spectral Graph Convolutions for Population-Based Disease Prediction

In the paper titled "Spectral Graph Convolutions for Population-based Disease Prediction," the authors, Parisot et al., propose an innovative application of Graph Convolutional Networks (GCNs) for disease prediction by leveraging both imaging and non-imaging data from large populations. The proposed method constructs sparse graphs where vertices represent subjects by imaging feature vectors, and edges encode pairwise phenotypic similarities. This approach addresses common limitations in modeling interactions in population data by integrating individual characteristics and contextual relationships.

Methodology

The authors develop an architecture using GCNs trained on partially labeled graphs. This framework aims to infer the labels of unlabeled nodes based on node features and pairwise associations. The graph framework introduced is twofold: it seeks to characterize individual sample features while simultaneously capturing global structural associations across the population using phenotypic data.

The implementation involves a structured approach:

1. Data Representation: Constructing the graph by representing subjects as nodes with associated imaging feature vectors and phenotypic attributes influencing edge weights.
2. Graph Convolutional Neural Networks: Employing spectral graph convolutions that extend traditional CNNs to process non-Euclidean data structures by leveraging graph signal processing techniques. Localized graph filters are defined in terms of Chebyshev polynomials to perform efficient spectral convolutions on graphs.

Applications and Results

The method was tested on two major datasets: the Autism Brain Imaging Data Exchange (ABIDE) and Alzheimer's Disease Neuroimaging Initiative (ADNI). The former entailed distinguishing controls from Autism Spectrum Disorder (ASD) patients, while the latter targeted predicting the conversion from Mild Cognitive Impairment (MCI) to Alzheimer's Disease (AD).

• ABIDE Dataset: The model achieved an accuracy of 69.5%, surpassing prior state-of-the-art results of 66.8%. This illustrates the method’s ability to account for heterogeneity in functional MRI data acquired from multiple sites.
• ADNI Dataset: The accuracy of prediction improved to 77% for MCI conversion, significantly outperforming standard linear models that only considered individual features. This enhancement demonstrates the robustness of integrating longitudinal data within the graph structure.

Implications and Future Directions

This research highlights the potential of GCNs in healthcare applications by effectively layering contextual inter-subject similarities alongside conventional diagnostic methods. It opens avenues for more personalized and accurate disease prediction algorithms.

Future work could focus on enhancing the complexity and expressiveness of feature vectors, potentially through autoencoders and deep learning-based embedding techniques. Moreover, incorporating attributed graphs where edges elucidate various dimensions of subject pairs can enrich the understanding of multifaceted population data. Additionally, temporal data integration may further enhance the model’s applicability to longitudinal studies, making it a promising direction for future inquiry.

The benefits of integrating phenotypic information with imaging feature vectors, as demonstrated in this study, underscore the importance of context-aware models in population-based health predictions. This innovative use of GCNs not only strengthens disease classification tasks but also paves the way for broader applications of graph-based learning methods in complex, large-scale biomedical data environments.