HyperGALE: ASD Classification via Hypergraph Gated Attention with Learnable Hyperedges

TL;DR

Evaluated on the extensive ABIDE II dataset, HyperGALE not only improves interpretability but also demonstrates statistically significant enhancements in key performance metrics compared to both previous baselines and the foundational hypergraph model.

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by varied social cognitive challenges and repetitive behavioral patterns. Identifying reliable brain imaging-based biomarkers for ASD has been a persistent challenge due to the spectrum's diverse symptomatology. Existing baselines in the field have made significant strides in this direction, yet there remains room for improvement in both performance and interpretability. We propose \emph{HyperGALE}, which builds upon the hypergraph by incorporating learned hyperedges and gated attention mechanisms. This approach has led to substantial improvements in the model's ability to interpret complex brain graph data, offering deeper insights into ASD biomarker characterization. Evaluated on the extensive ABIDE II dataset, \emph{HyperGALE} not only improves interpretability but also demonstrates statistically significant enhancements in key performance metrics compared to both previous baselines and the foundational hypergraph model. The advancement \emph{HyperGALE} brings to ASD research highlights the potential of sophisticated graph-based techniques in neurodevelopmental studies. The source code and implementation instructions are available at GitHub:https://github.com/mehular0ra/HyperGALE.

Figures (7)

• Figure 1: ASD Processing Pipeline and the Proposed HyperGale Architecuture. ( a) fMRI data is preprocessed with Schaefer's parcellation and later converted into a functional connectivity (FC) matrix. (b) The FC matrix is then converted into a hypergraph with learnable hyperedge weights. Subsequently, the hypergraph features are sent to a series of hypergraph convolution layers and gated attention is applied in order to capture the importance of relevant information. Finally, the information extracted after attention layers is aggregated using the readout function. Finally, from these lower dimensional features of readout, a sigmoid activation leads to binary classification of ASD vs Typically Developing (TD).
• Figure 2: Modified hypergraph convolution proposed in HyperGALE pipeline. Starting with ROI features derived from the fully connected matrix, these initial inputs undergo node-to-hyperedge propagation to form hyperedge features (step (b)). These features are scaled using learned weights, as in step (c), the activation happens in different hyperedges with different values. Finally, the scaled hyperedge features are propagated to new node features (step (d)).
• Figure 3: Illustration of the Gated Attention Module. GA module which learns $\alpha$ iteratively (or numerically) from eq. (\ref{['eq-hypergraph-gated-attention']}), this $\alpha$ vector is multiplied to get the final node features which is fed to the readout layer. ilse2018attention.
• Figure 4: Prediction performance across sites using a leave-one-site-out strategy. The count of ASD and TD subjects from various sites are shown on (left) and various performance metrics are shown for each site on (right). Chance level performance is at 50%. Despite the challenges with site-specific variations in the number of samples, our model is still able to demonstrate creditable between-site generalization performance, comparable to the results in Table \ref{['tab-performance']}.
• Figure 5: Changes in prediction accuracy with respect to change in number of ROIs in a hyperedge denoted by k. The optimal number of ROIs was found at $k$ = 40.