AMGL: Adaptive Multimodal Graph Learning for Brain Disease Prediction

Graph-based approaches have been widely adopted in biomedical applications for modeling multimodal data, particularly in the accurate diagnosis and effective treatment of brain disorders. Most existing graph-based multimodal medical data processing methods typically extract features by fusing multimodal information through weighted operations, and then manually define graph structures based on specific modalities to learn patient representations via graph embedding. However, these methods often overlook the complex correlations and discrepancies across modalities, making it difficult to obtain highly relevant information. Moreover, the prior construction of an appropriate graph presents a considerable challenge, as manually defined structures are susceptible to spurious or noisy edges. These factors inevitably lead to incorrect predictions in real-world clinical scenarios. To address these limitations, we propose an end-to-end adaptive multimodal graph learning (AMGL) framework that comprises two key modules: modal-aware integration learning (MAIL) and cluster-constrained adaptive graph learning (CAGL). MAIL captures both intermodal relevance and complementarity to construct enriched modality-aware representations, while CAGL performs adaptive graph learning based on data clustering and utilizes a graph-gated neural network (GGNN) for disease prediction. The experimental results on the TADPOLE and autism brain imaging data exchange (ABIDE) datasets demonstrate that our method achieves superior classification accuracy and generalization capability, with an average performance gain of 2%–3% over state-of-the-art approaches. The code of AMGL is available at https://github.com/WuJUNEhhh/AMGL