CTAD (2025)
Juhyun Lee, Sungjoon Park, Kyungwook Lee, Soorin Yim, Doyeong Hwang, Kiwoong Yoo, Soonyoung Lee, Kiyoung Kim
Abstract
Background
Recent advances in machine learning have produced predictive models that, in some cases, surpass physicians and pathologists in disease classification. In Alzheimer’s Disease (AD), a highly complex brain disorder, AI-based approaches are now used to predict disease status and progression from multimodal data, while also identifying key predictive features. Clinically, AD diagnosis draws from diverse sources: cognitive assessments, medical history, genetics, and neuroimaging [1]. Integrating such multimodal information offers a more holistic view of disease trajectory.
However, many current studies are limited by requiring complete datasets, excluding participants missing data in any modality. This constraint is particularly problematic in real-world research and clinical trial settings where incomplete data is common, leading to the exclusion of valuable patient records.
We address this limitation by introducing MOIRA (Multimodal Integration with Robustness to Absent modalities), a framework specifically designed to handle incomplete heterogeneous data. By leveraging available modalities while compensating for missing ones, MOIRA increases data coverage without compromising predictive performance. This is especially critical for AD research and clinical applications, as it enables inference even with partial records and broadens participant inclusion.
Methods
We evaluated MOIRA on two widely used AD datasets. The first, Religious Order Study and Memory and Aging Project (ROSMAP) [2], contains five post-mortem omics modalities: mRNA expression, DNA methylation, microRNA expression, tandem mass tag proteomics, and HD4 metabolite quantification. Only 86 of 748 samples (11%) had all five modalities; most were missing at least one.
The second dataset, the Alzheimer’s Disease Neuroimaging Initiative (ADNI) [3], was organized into four modalities: Imaging, Clinical, Biospecimen, and Genomics. The Imaging modality comprised 3D T1-weighted MRI scans processed through a standardized pipeline with FSL library. The Clinical modality included demographics and medical history. The Biospecimen modality encompassed CSF biomarkers— amyloid-β, total tau, and phosphorylated tau—established measures for diagnosis and clinical-trial enrichment in AD [4], as well as the ApoE genotype, a major genetic risk factor frequently considered in trial design and drug development [5]. Genomic data consisted of single nucleotide polymorphisms (SNPs) in PLINK format. Of the 2,380 participants, only 1,323 had complete data across all four modalities. MOIRA was designed to utilize incomplete cases, thereby expanding cohort coverage by approximately 80%.
The architecture comprises three components: (1) modality-specific encoders that embed each modality into a latent representation, (2) a fusion module that integrates modality embeddings with learnable weights and redistributes contributions to present modalities by masking absent inputs, and (3) a predictor for classification.
Feature selection reduced dimensionality while preserving signal. We used ANOVA F-tests for ROSMAP omics and ADNI Genomics. For ADNI Imaging, MRI scans were processed using a 3D convolutional neural network with three convolutional layers (16, 32, and 64 channels), LeakyReLU activations, 3D max pooling, and a final 128-dimensional flattened embedding. Other modalities were passed into encoders without additional preprocessing.
The training objective combined prediction loss (cross-entropy on final outputs), auxiliary losses (cross-entropy on modality-specific embeddings to preserve discriminative capacity), and a contrastive loss (CLIP-style across modality pairs) to encourage alignment in the shared embedding space.
Results
In ROSMAP, we conducted binary classification of cognitive diagnosis. MOIRA achieved an accuracy of 0.92, significantly outperforming previous methods. This gain was largely due to its ability to incorporate incomplete cases: earlier models were restricted to 391 fully observed samples, while MOIRA utilized 784 participants.
Ablation studies confirmed that including all five modalities improved accuracy over three-modality models, and that union-based training (using all available samples) outperformed intersection-only approaches. These results demonstrate the clinical value of retaining partially observed data.
In ADNI, MOIRA attained 0.91 accuracy in three-way classification among AD, mild cognitive impairment (MCI), and normal controls (NC). Other models that tolerate missing modalities often suffer reduced accuracy with larger, more incomplete cohorts. In contrast, MOIRA increased coverage from 1,323 to 2,380 participants while maintaining high performance—the best trade-off between accuracy and cohort size among models handling incomplete data. This is especially valuable for clinical trials, where maximizing patient inclusion without sacrificing diagnostic fidelity is essential.
Conclusion
We present MOIRA, a robust multimodal integration framework that enables meaningful analysis of AD data even when modalities are missing. The inference under partial modality availability greatly expands usable cohort size without compromising prediction accuracy. In both ROSMAP and ADNI, it demonstrated that incomplete records can be leveraged effectively, increasing statistical power and avoiding unnecessary patient exclusion. By overcoming the constraints of requiring complete datasets, MOIRA makes predictive modeling more applicable to real-world clinical practice and research. Our findings underscore the importance of embracing incomplete yet informative data in AD studies. Our results suggest that future clinical studies and interventional trials in AD can benefit from adopting methods that explicitly accommodate missing modalities, improving both coverage and reliability. By bridging the gap between idealized complete datasets and the realities of clinical data collection, MOIRA offers a pathway toward more inclusive, accurate, and clinically meaningful applications of AI in AD.