Contextualized biomedical language processing enhances ICU survival prediction.

Area of Science:

• Biomedical informatics
• Clinical data science
• Natural language processing in healthcare

Background:

• Accurate prediction of intensive care unit (ICU) survival is difficult due to complex and varied patient data.
• Existing models often struggle to fully leverage unstructured clinical notes and diagnostic codes.

Purpose of the Study:

• To investigate the impact of contextualized biomedical language processing on ICU survival prediction.
• To develop and validate multimodal models that integrate structured and unstructured clinical data.

Main Methods:

• Trained and validated multimodal models using BioBERT-enhanced convolutional neural networks on MIMIC-IV, MIMIC-III, and eICU datasets.
• Integrated structured laboratory data with unstructured text, including chief complaints and International Classification of Diseases (ICD) entries.
• Evaluated model performance using area under the receiver operating characteristic curves (AUROCs).

Main Results:

• The BioBERT-enhanced model achieved high AUROCs (0.889 strict, 0.974 lenient cohort) during external validation.
• Excluding text features or using coded ICD entries reduced predictive performance, underscoring the value of contextual embeddings.
• A secondary task of cerebrospinal fluid culture prediction achieved an AUROC of 0.853.

Conclusions:

• Integrating contextualized biomedical language representations significantly enhances multimodal learning for ICU survival prediction.
• Unstructured clinical text, when processed with advanced language models, provides crucial information for improving patient outcome prediction.
• This approach offers a promising direction for advancing clinical decision support systems in critical care.