HDGL: A hierarchical dynamic graph representation learning model for brain disorder classification

TL;DR

HDGL presents a novel hierarchical dynamic graph learning framework that jointly models dynamic brain graphs and a population graph to classify brain disorders from rs-fMRI data. It integrates a sliding-window-based dynamic FC, GRU-based node features, GCN/SAGPool, and a Transformer to obtain robust brain-graph embeddings, followed by a Graph Transformer on a phenotype-informed population graph. The model explores four training schemes (three transductive, one inductive) to balance performance and memory. Across ABIDE and ADHD-200, HDGL variants outperform baselines, with ablations highlighting the importance of cross-subject relationships, temporal modeling, and phenotype-informed population graphs. This approach advances brain disorder diagnosis by combining dynamic connectivity, hierarchical graph structure, and multimodal information into an end-to-end learnable pipeline with practical scalability.

Abstract

The human brain can be considered as complex networks, composed of various regions that continuously exchange their information with each other, forming the brain network graph, from which nodes and edges are extracted using resting-state functional magnetic resonance imaging (rs-fMRI). Therefore, this graph can potentially depict abnormal patterns that have emerged under the influence of brain disorders. So far, numerous studies have attempted to find embeddings for brain network graphs and subsequently classify samples with brain disorders from healthy ones, which include limitations such as: not considering the relationship between samples, not utilizing phenotype information, lack of temporal analysis, using static functional connectivity (FC) instead of dynamic ones and using a fixed graph structure. We propose a hierarchical dynamic graph representation learning (HDGL) model, which is the first model designed to address all the aforementioned challenges. HDGL consists of two levels, where at the first level, it constructs brain network graphs and learns their spatial and temporal embeddings, and at the second level, it forms population graphs and performs classification after embedding learning. Furthermore, based on how these two levels are trained, four methods have been introduced, some of which are suggested for reducing memory complexity. We evaluated the performance of the proposed model on the ABIDE and ADHD-200 datasets, and the results indicate the improvement of this model compared to several state-of-the-art models in terms of various evaluation metrics.

Figures (13)

• Figure 1: Illustration of various types of graphs defined in brain disorder classification methods, assuming we have $m$ subjects and $N$ ROIs: (a) Brain-network based approaches (b) population-graph based approaches (c) hierarchical-graph based approaches
• Figure 2: HDGL architecture consists of four parts: (a) Adjacency matrix construction (b) Node feature extraction (c) Brain network analysis (d) Population network analysis. From another perspective, HDGL consists of two levels: the Brain network-level, which constructs the brain network graphs and learns embeddings for them, and the Population network-level, which constructs the population graph, learns embeddings for its nodes, and then classifies them.
• Figure 3: (a) The dynamic brain graph obtained for the $i$-th subject under the assumption of $T$=3. (b) Adjacency matrix of dynamic brain graph.
• Figure 4: The dynamic brain graph and its associated adjacency matrix, before and after passing through the SAGPool layer, where $N$ is the number of ROIs and $k \in [0, 1)$ indicates the pooling ratio.
• Figure 5: The general outline of part (c) for $K$=3.