Explainable multiplex graph propagational network with multimodal neuroimage integration for dementia subtype diagnosis.

Abstract

Dementia encompasses diverse subtypes with distinct characteristics, including cognitive functions, cerebrospinal fluid biomarkers, and neuroimages. Neuroimaging-based diagnosis is advantageous due to low variability and minimal invasiveness, ensuring safe and accurate outcomes. Recently, integrating multimodal neuroimages with machine learning techniques has enhanced diagnostic precision. Especially, graph neural network (GNN) has emerged as promising models by considering connectivity between brain regions. However, current GNN-based methods are limited by focusing solely on local connectivity, failing to adequately capture global interactions crucial to structural pathways and functional brain activities. Additionally, existing methods pose a trade-off between improving diagnostic performance and maintaining explainability, as complex feature transformations across hidden layers obscure model explanations. This limitation is especially critical in clinical settings, where transparent decision-making directly impacts patient outcomes. Motivated by these limitations, we propose a novel method for diagnosing dementia subtypes by integrating multimodal neuroimages, called Explainable Multiplex Graph Propagational Network (EMGPN). Our method employs multiplex graphs derived from multiple neuroimaging modalities, propagating features across brain regions to concurrently represent local and global connectivity. EMGPN subsequently integrates these multimodal features through region-specific, probabilistically derived parameters, thus preserving individual modality characteristics. Crucially, EMGPN maintains explainability through a transparent architecture without hidden layers, allowing clinicians to clearly understand model outcomes. We validated EMGPN using an elderly South Korean cohort across various dementia subtypes. The results indicated that EMGPN achieved an average performance improvement of 8.6 % compared to existing methods, while generating explainable outputs, including region-specific modality contributions and subtype-specific brain region importance maps. These findings underscore EMGPN's significant potential as a clinically applicable, explainable, and robust tool for neuroimaging-based dementia diagnosis.