ICU Mortality Prediction via Multimodal Learning on MIMIC-IV Clinical Data

Purpose: This study proposes a multimodal framework for retrospective short-term mortality prediction in intensive care unit patients by integrating structured admission data with semantic information extracted from discharge summaries. The objective is to improve patient characterization by combining complementary clinical data that are often analyzed separately. Methods: Structured inputs, including demographics, administrative records, laboratory results, and clinical severity scores, were merged with semantic representations of clinical notes produced by a BERT-based encoder. The resulting embeddings were reduced in dimensionality and clustered with K-Means to obtain categorical semantic features reflecting thematic similarity among notes. These semantic cluster identifiers were then concatenated with tabular variables to create a unified multimodal representation, which was used to train five classical machine learning classifiers. Results: The integration of semantic features consistently improved predictive performance compared with models trained only on structured data. Among the evaluated algorithms, Random Forest achieved the best discrimination, with an area under the receiver operating characteristic curve of 0.940 in the multimodal setting. Conclusion: Combining structured clinical variables with unstructured admission notes enhances short-term mortality prediction and supports the development of interpretable multimodal decision-support tools for intensive care settings.