LG AI Research: MOIRA, an Alzheimer's Prediction Model That Overcomes Incomplete Multimodal Data

In real clinical settings, it is difficult for every patient to complete all tests—MRI, genomics, biomarkers, etc.—due to cost or invasiveness issues. Existing multimodal AI research used only 'complete data' with all items present, resulting in limited data utilization and reduced generalization performance.

MOIRA (Multi-Omics Integration with Robustness to Absent Modalities), developed by LG AI Research, is a framework that provides reliable predictions even when modalities are missing. This model consists of three key components.

• Encoders: Map data from different modalities into an embedding space of the same dimension
• Adaptive Fusion Module: Assigns weights only to existing modalities and masks missing data to generate an integrated embedding
• Predictor: Performs final classification based on the integrated embedding

During training, prediction loss, auxiliary embedding loss, and a CLIP-style Contrastive Loss are combined to optimize alignment across modalities and prevent Modality Collapse.

Experimental results show that MOIRA achieved accuracy of 0.920 on ROSMAP and 0.911 on ADNI, standard datasets in Alzheimer's research, attaining state-of-the-art performance compared to existing research. In particular, it resolves the trade-off problem of existing models where accuracy drops as data volume increases, securing both dataset scalability and prediction accuracy simultaneously.