Geometry-guided unimodal-to-multimodal composition for incomplete and asynchronous clinical information fusion

Clinical prediction often requires fusing heterogeneous data sources such as coded patient context, lab or vital trajectories, and free-text documentation, but the set of observable modalities varies across care workflows and observation timepoints. We study this setting as an incomplete clinical information fusion problem rather than as explicit longitudinal sequence modeling. ETF-UML addresses the challenge by training modality-specific encoders independently while aligning them to a shared Equiangular Tight Frame (ETF) decision geometry, then composing only the observed modalities through metadata-conditioned fusion. This design targets robust unimodal-to-multimodal composition when modality acquisition is asynchronous, incomplete, and time-varying. We clarify the training and evaluation protocol and add representation analysis to illustrate how the shared geometry structures unimodal and fused spaces. We report the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), and their average. On MIMIC-IV 30/90-day mortality and eICU 10-day in-hospital mortality, the reported aggregate results indicate consistent gains in AUROC and in average performance over end-to-end fusion and missing-modality baselines. Additional experiments attribute the gains to the sharing and the fixing of the decision geometry, and show that the advantage holds on a task with more than two classes and under reduced modality availability.