Deep Phenotyping of Obstructive Sleep Apnea Using Multimodal Representation Learning

Obstructive sleep apnea (OSA) is linked to elevated cardiovascular and metabolic risks. We developed a multimodal framework integrating structured clinical data and unstructured discharge summaries using SHAP-optimized TabNet and multiple clinical text encoders with late fusion to generate unified patient embeddings. After UMAP dimensionality reduction, multimodal clustering produced more cohesive and better-separated structures than single-modality approaches (Silhouette Score: 0.56 vs. ~0.4) and revealed multiple clinically distinct OSA subgroups (seven after survival-based merging) with marked differences in comorbidity profiles and survival (p < 0.0001). Multimodal classification achieved improved AUROC for phenotype prediction across encoder configurations, with the strongest performance observed using GatorTron-based fusion (AUROC: 0.918 vs. 0.894 structured baseline). These results demonstrate that integrating heterogeneous data improves OSA patient stratification and phenotype prediction, and that encoder choice plays a critical role in multimodal representation quality.