Propensity-guided unbalanced optimal transport for acquisition-aware incomplete multimodal clinical prediction

Multimodal learning is important for clinical decision-making, yet clinically indicated modalities such as chest X-rays (CXR) are often unavailable in real-world datasets. Since CXR acquisition is driven by clinical need and physician ordering decisions, this selective modality acquisition pattern is clinically informative and can lead to acquisition-related incomplete modality availability. Existing missing-modality methods mainly address incomplete inputs through adaptive fusion or generative imputation, but rarely model the acquisition-induced distribution discrepancy between the CXR-observed cohort and the broader patient population. We propose PropUOT, an acquisition-aware distribution alignment framework for multimodal clinical prediction. PropUOT estimates CXR acquisition propensity from source information, performs bounded propensity-guided latent adjustment for CXR-missing samples, and uses Unbalanced Optimal Transport (UOT) to regularize the adjusted source-side representations with observed CXR representations under relaxed mass constraints. This avoids imputing missing CXR features while mitigating acquisition-induced representation discrepancy. Experiments on public Medical Information Mart for Intensive Care (MIMIC) datasets demonstrate that PropUOT achieves state-of-the-art performance under both partially matched and fully matched multimodal clinical prediction settings.