ProgMF: Progressive Multimodal Representation Learning for Incomplete Biomedical Data

Multimodal learning in biomedical research is often limited by unequal modality coverage, particularly in Alzheimer's disease (AD) neuroimaging, where acquisition costs and limited access to specialized imaging reduce the number of subjects observed across all modalities. Conventional complete-view fusion methods exclude partially observed subjects, while imputation methods reconstruct missing observations and may introduce reconstruction errors. We introduce Progressive Multimodal Fusion (ProgMF), a framework that learns from larger cohorts with fewer modalities and transfers the learned representations as additional modalities are incorporated. Using independent vector analysis (IVA) as a testbed, ProgMF independently learns a joint representation at each stage and uses the preceding stage's representation as an anchor to guide subsequent refinement. This enables knowledge transfer without imputing missing observations while allowing adaptation to dependencies introduced by new modalities. Controlled simulations involving four modalities show that ProgMF achieves lower latent source recovery error than complete-view fusion and imputation-based baselines. A controlled ablation further demonstrates the contribution of subjects available only at earlier stages. Evaluation on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset also shows higher accuracy and F1 Score than the evaluated baselines for cognitively normal versus dementia classification, providing preliminary evidence of improved generalization to held-out subjects.