Dynamic causal graph representation learning for neurological disorders recognition

Abstract

The effectiveness of dynamic brain networks in diagnosing neurological disorders, which represent a worldwide threat to human health, is increasingly recognized. Numerous Graph Neural Network (GNN) models have been created to autonomously detect aberrant integration patterns related to specific neurological conditions in brain graphs. Nevertheless, most existing approaches exhibit primary drawbacks. These methodologies typically either employ predetermined statistical metrics (e.g., Pearson correlation) or depend on static brain networks. Moreover, the process of constructing dynamic brain networks/graphs concentrates on static or temporally invariant effective connectivity, resulting in a substantial knowledge gap about the dynamic aspect of dEC. Although statistically-specific brain networks are extremely efficient in identifying certain disorders, their performance may lack robustness when applied to other neurological diseases. To address these issues, we introduce a novel model termed the Dynamic Causal Graph Inference Graph Neural Network (DCGI-GNN), which can estimate brain networks and learn graph representation within a two-stage pipeline, facilitating the optimization and interpretability of each stage. The term “causal” in this work refers to directed, time-varying functional influence relationships, not intervention-based causality. Our model initially deduces the task-relevant graph structures directly from the data and later utilizes them to improve downstream task performance. We conducted extensive experiments to diagnose three brain diseases: autism spectrum disorder (ASD), Alzheimer’s disease (AD), and Parkinson’s disease (PD) using data from the ABIDE, ADNI, and PDD datasets. Our findings demonstrate the effectiveness of the proposed framework. Beyond its superior performance, the model is highly interpretable, and its conclusions align largely with existing literature.