[CTAD 2025] MOIRA, a Multimodal Framework That Predicts Alzheimer's More Effectively with Incomplete Data

To predict complex brain diseases like Alzheimer's, a Multimodal approach that integrates diverse data such as genomics, MRI, and cognitive tests is essential. However, in real clinical settings, the Incomplete modality problem frequently occurs, where certain test data is missing due to cost or invasiveness issues.

Existing multimodal AI models used only samples with fully complete data for training, so for patients with partially missing data, the model's prediction accuracy would drop or such patients would be excluded from analysis.

MOIRA (Multi-Omics Integration with Robustness to Absent modalities), developed by LG AI Research, is a framework designed to solve this problem. MOIRA maintains high prediction accuracy even when data is missing, and has the following key structure.

• Encoders: Maps different types of data into an embedding space of the same dimension through dedicated encoders for each modality.
• Adaptive Fusion Module: Integrates features from different modalities to generate a combined representation.
• Predictor: Performs the final disease prediction based on the integrated embedding.

Through this technology, the research team laid the groundwork for effectively utilizing patient data without exclusion, even in situations where key data such as MRI or CSF (cerebrospinal fluid) biomarkers is missing.