Structural MRI–Informed Multimodal Fusion for Robust Alzheimer’s Disease Prediction

Juhyun Lee, Sungjoon Park, Kyungwook Lee, Soorin Yim, Doyeong Hwang, Dongyun Kim, Kiwoong Yoo, Soonyoung Lee, Kiyoung Kim

Submission Track: Short papers presenting ongoing research or work submitted to other venues (up to 5 pages, excluding references)

Keywords: Multimodal, Multi-omics Integration, ImageOmics, Alzheimer’s Disease Prediction, Representation Learning, Biomedical AI

TL;DR: MOIRA, a framework that integrates sMRI and omics for AD prediction, remains flexible to incomplete modalities and outperforms prior approaches on the ADNI dataset.

Abstract: Alzheimer’s Disease (AD) is a devastating neurodegenerative disorder, and accurate prediction remains a critical challenge. Multiple modalities can inform AD prediction, with public resources such as the Alzheimer's Disease Neuroimaging Initiative (ADNI) providing access to diverse datasets. However, prior studies often rely on single modalities, limiting clinical applicability, or struggle to integrate multimodal data effectively. In this work, we introduce MOIRA (Multi-Omics Integration with Robustness to Absent modalities), a predictive framework that leverages the strong discriminative power of structural Magnetic Resonance Imaging (sMRI) while flexibly incorporating additional modalities to boost performance. MOIRA achieves 0.91 accuracy, substantially surpassing existing approaches. Notably, we show that our model trained with sMRI can still improve prediction without sMRI data at inference, supporting potentially cost-efficient diagnostic strategies in clinical settings. Our findings highlight the value of sMRI-informed multimodal integration for advancing robust, translational AD prediction.

Submission Number: 5