Taming arbitrary modality missingness and imbalance: A unified graph-MoE framework for Alzheimer's disease diagnosis

Multimodal biomarkers hold significant potential for improving Alzheimer's disease (AD) diagnosis. However, existing multimodal learning methods typically rely on idealized assumptions of complete and balanced modality availability, which rarely hold in clinical practice. Real-world datasets are plagued by arbitrary modality missingness and, more critically, a long-tailed modality-combination imbalance stemming from hierarchical clinical protocols. These issues cause conventional models to be biased toward frequent combinations while failing on rare but diagnostically critical ones for early AD diagnosis. To address these challenges, we propose UniMIX-AD, a unified framework tailored for robust AD diagnosis under realistic, imperfect multimodal settings. UniMIX-AD integrates two core innovations: (i) a Unified Missing-modality Imputation (UMI) module that dynamically reconstructs representations of missing modalities by capturing cross-modal dependencies from any observed modality combination, ensuring comprehensive patient representations; and (ii) a Graph-coordinated Sparse Mixture-of-Experts (G-Sparse-MoE) that adaptively routes patients to specialized experts and facilitates structured knowledge transfer from frequent (head) to rare (tail) combinations, effectively alleviating the long-tailed modality-combination imbalance. Extensive experiments on the ADNI and OASIS-3 datasets demonstrate that UniMIX-AD consistently outperforms state-of-the-art methods, achieving a relative average accuracy improvement of 7.1% and a significant boost of up to 14.3% on tail modality combinations in ADNI.