Overview
This paper presents a novel approach to predicting extubation failure (EF) by leveraging features extracted from free-text respiratory therapy clinical notes using a large language model (LLM) paired with a logistic regression pipeline.
Background
Invasive mechanical ventilation is a lifesaving therapy, but timely and safe discontinuation is essential to preventing extubation failure and its associated health risks. Accurate EF prediction remains a critical challenge in intensive care, where both premature and delayed extubation can lead to adverse patient outcomes.
Methodology
The authors apply their LLM-based feature classification approach to a patient cohort from University of Washington Medicine. The pipeline identifies clinically meaningful EF-related features from unstructured respiratory therapy notes, which are then integrated with structured patient data to improve prediction performance.
Key Contributions
- A novel EF prediction framework that combines LLM-derived features from free-text clinical notes with conventional structured data.
- Identification of clinically meaningful EF-related features that measurably improve prediction performance when included alongside structured patient data.
- A critical analysis of how differences in target populations across prior EF prediction studies—such as heterogeneous inclusion criteria and varying EF definitions—can lead to systematic differences in model performance and hinder generalizability between studies.
Publication Details
- Venue: Published in CHIL 2026 (Conference on Health, Inference, and Learning)
- Length: 11 pages, 4 figures, 4 tables (25 pages including citations and supplemental material)
- Authors: Izzy Chaiken, Aditya Khowal, Neha A. Sathe, Mark M. Wurfel, Lucy Lu Wang
- Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG); Physics and Society (physics.soc-ph)
- ACM Classes: I.2.7; J.3
- arXiv: arXiv:2609.17532 [cs.CL]
- DOI: https://doi.org/10.48550/arXiv.2609.17532
- Submitted: 17 June 2026
Significance
As LLMs continue to mature in 2026, this work demonstrates a practical pathway for translating their text-understanding capabilities into clinically actionable features. By bridging unstructured clinical documentation with predictive modeling, the approach addresses a longstanding gap in critical care informatics: the wealth of signal contained in respiratory therapy notes that traditional structured-data models overlook. The paper's emphasis on cross-study comparability also offers a methodological caution for the broader clinical ML community, where inconsistent outcome definitions frequently undermine reproducibility and generalization.
via ArXiv CL+LG