MREF-AD: New Mixture-of-Experts model for interpretable Alzheimer's diagnosis

Accurate early diagnosis of Alzheimer's disease (AD) requires integrating complementary information from multimodal neuroimaging, such as amyloid PET and MRI. Traditional fusion methods often naively concatenate features, failing to adaptively balance contributions from different biomarkers and brain regions. A new model addresses this gap with interpretability at its core.

MREF-AD (Multimodal Regional Expert Fusion) employs a Mixture-of-Experts (MoE) framework that treats mesoscopic brain regions within each modality as independent experts. A learned gating network dynamically weights these experts for each subject, producing subject-specific fusion weights. Using tabular neuroimaging and demographic data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), MREF-AD achieves competitive performance against both classic and deep learning baselines.

Unlike black-box approaches, MREF-AD provides interpretable insights at the modality and region level—showing exactly how structural and molecular imaging contribute to the diagnosis. This transparency is crucial for clinical trust. The work was published at IEEE ICHI 2026 (arXiv:2512.10966) and includes open-source code.

For practitioners, MREF-AD demonstrates that specialized regional experts with adaptive weighting can outperform simple feature concatenation, offering a blueprint for interpretable multimodal AI in healthcare. The model's ability to highlight which brain regions and imaging modalities drive each diagnosis could help clinicians validate findings and personalize treatment plans.