Heterogeneous Graph Convolutional Neural Network via Hodge-Laplacian for Brain Functional Data.

Area of Science:

• Neuroscience
• Machine Learning
• Graph Neural Networks

Background:

• Functional brain connectivity analysis using fMRI data is complex.
• Existing graph neural networks (GNNs) face challenges in handling heterogeneous brain data.
• Accurate prediction of cognitive abilities like general intelligence from brain activity remains a challenge.

Purpose of the Study:

• To propose a novel heterogeneous graph convolutional neural network (HGCNN) for analyzing brain fMRI data.
• To introduce a generic formulation of spectral filters using the k-th Hodge-Laplacian (HL) operator and a topological graph pooling (TGPool) method.
• To evaluate the performance of HL-node, HL-edge, and HL-HGCNN models in predicting general intelligence from fMRI data.

Main Methods:

• Developed a novel heterogeneous graph convolutional neural network (HGCNN) utilizing the k-th Hodge-Laplacian (HL) operator for spectral filtering.
• Introduced a topological graph pooling (TGPool) method applicable to simplex graphs of any dimension.
• Designed and implemented HL-node, HL-edge, and HL-HGCNN architectures for learning signal representations at node, edge, and combined levels.
• Utilized fMRI data from the Adolescent Brain Cognitive Development (ABCD) study (n=7693) to predict general intelligence.

Main Results:

• The HL-edge network demonstrated superior performance compared to the HL-node network when functional brain connectivity was used as features.
• The proposed HL-HGCNN significantly outperformed state-of-the-art GNNs, including GAT, BrainGNN, dGCN, BrainNetCNN, and Hypergraph NN.
• Functional connectivity features extracted by the HL-HGCNN were found to be meaningful for interpreting neural circuits associated with general intelligence.

Conclusions:

• The novel HGCNN framework, incorporating HL spectral filters and TGPool, provides an effective approach for analyzing complex brain fMRI data.
• The HL-HGCNN model shows significant potential for predicting cognitive abilities and understanding the underlying neural mechanisms.
• Considering functional brain connectivity at the edge level enhances the predictive power of graph-based neural networks for intelligence prediction.