Contextualized biomedical language processing enhances ICU survival prediction

Accurate prediction of intensive care unit (ICU) survival remains challenging due to heterogeneous clinical data. This study shows that contextualized biomedical language processing markedly enhances ICU survival prediction. Multimodal models integrating structured laboratory data with unstructured text (chief complaints and International Classification of Diseases [ICD] entries) were trained and validated using MIMIC-IV, MIMIC-III, and eICU datasets. The BioBERT-enhanced convolutional neural network achieved area under the receiver operating characteristic curves (AUROCs) of 0.889 (strict cohort, <i>n</i> = 5,795) and 0.974 (lenient cohort, <i>n</i> = 58,615) during external validation. Excluding text features or replacing free-text ICD entries with coded formats reduced performance (AUROC from 0.983 to 0.946-0.947), highlighting the importance of contextual embeddings. As a secondary task, cerebrospinal fluid culture prediction achieved AUROC = 0.853. Overall, integrating contextualized biomedical language representations significantly improves multimodal learning and ICU survival prediction.