LLM-derived features from respiratory therapy notes improve extubation failure prediction in clinical study

A preprint from researchers at the University of Washington Medicine introduces a pipeline that uses a large language model to classify free-text respiratory therapy notes for features relevant to extubation failure (EF), a complication where patients require reintubation after removal of a breathing tube.

The team reports that including LLM-derived features alongside structured patient data improved EF prediction performance in a cohort from University of Washington Medicine. The study emphasizes that prior EF prediction work often suffers from heterogeneous inclusion criteria and inconsistent EF definitions, which can systematically affect model performance and limit generalizability across settings.

The authors propose their method as a way to leverage unstructured clinical text—common in electronic health records—to extract signals that structured data alone may miss, while cautioning that differences in patient populations and outcome definitions complicate cross-study comparisons.

The paper is titled “Enhancing Extubation Failure Prediction with LLM-Derived Features from Respiratory Therapy Clinical Notes” and was submitted to the Conference on Health, Inference, and Learning (CHIL) 2026.