A multicenter study has demonstrated that an interpretable machine learning model utilizing data collected exclusively during early emergency department resuscitation can accurately predict extubation failure in mechanically ventilated patients experiencing an acute exacerbation of chronic obstructive pulmonary disease. The random forest classifier achieved an area under the receiver operating characteristic curve of 0.880 in an internal validation cohort and 0.874 in a prospective external validation cohort, providing clinicians with upstream risk identification to guide intensive care unit management.
Weaning intubated patients with acute exacerbations of chronic obstructive pulmonary disease from mechanical ventilation remains a primary challenge in critical care medicine. Existing pre-extubation assessments frequently exhibit suboptimal accuracy when evaluating weaning readiness. While machine learning algorithms have shown predictive promise, previously developed models have been constrained by opaque black-box architectures and an absence of rigorous external validation.
To develop a transparent predictive tool, researchers conducted a retrospective multicenter investigation enrolling 535 consecutive mechanically ventilated patients with acute exacerbations of chronic obstructive pulmonary disease at Yixing People’s Hospital between January 2018 and January 2023. This initial population was randomly partitioned at a seven-to-three ratio using outcome-stratified sampling into a derivation set of 375 patients and an internal validation set of 160 patients. For temporal external validation, investigators established a prospective observational cohort of 200 patients at Jiangsu Provincial People’s Hospital between May 2023 and May 2025.
Investigators gathered 19 pre-intubation variables and implemented a dual-stage LASSO-Boruta feature selection pipeline to identify robust predictors. Six distinct machine learning algorithms were trained with class-weight adjustments to address outcome imbalance. The team evaluated discriminatory performance using the area under the receiver operating characteristic curve, calibration curves, and decision curve analysis, integrating the SHapley Additive exPlanations framework to ensure algorithmic interpretability.
The analysis identified seven core predictive variables: the oxygenation index, left ventricular ejection fraction, white blood cell count, baseline lactate, diaphragmatic excursion, chronological age, and serum creatinine. The random forest model delivered optimal discrimination, recording area under the curve values of 0.880 with a 95 percent confidence interval of 0.814 to 0.947 internally, and 0.874 with a 95 percent confidence interval of 0.803 to 0.945 externally. Decision curve analysis confirmed net clinical benefit across relevant risk thresholds, with interpretability profiling showing that early baseline oxygenation, intrinsic cardiac reserve, and systemic inflammatory burden were the primary drivers in a newly developed web application for real-time risk stratification.
The incorporation of the SHapley Additive exPlanations framework specifically addressed the opacity that has historically limited machine learning applications in critical care. By examining how individual variables shaped predictions, the analysis revealed that early baseline oxygenation status measured by the oxygenation index, intrinsic cardiac reserve reflected by left ventricular ejection fraction, and systemic inflammatory burden indicated by white blood cell counts served as the primary drivers of extubation failure risk. These primary physiological factors operated in conjunction with baseline lactate, diaphragmatic excursion, chronological age, and serum creatinine to form a transparent multimodal predictive profile.
The modeling strategy also specifically confronted the challenge of outcome imbalance encountered in clinical datasets. By applying class-weight adjustments across all six machine learning algorithms, the investigators accounted for distribution discrepancies during model training. In addition, the dual-stage LASSO-Boruta pipeline systematically screened the nineteen pre-intubation variables collected during early emergency department resuscitation to isolate the seven most robust predictors, ensuring that the resulting random forest model avoided unnecessary complexity while capturing critical early resuscitation signals.
Rigorous validation across separate patient cohorts demonstrated the stability and clinical relevance of the chosen architecture. Following training and testing in the derivation and internal validation cohorts at Yixing People’s Hospital, the random forest classifier maintained consistent predictive accuracy when evaluated in the prospective temporal cohort at Jiangsu Provincial People’s Hospital. Calibration curves confirmed robust calibration between predicted probabilities and observed outcomes, while decision curve analysis verified favorable net clinical benefit across relevant clinical risk thresholds, underscoring the model’s reliability across different hospital environments.
To translate these computational findings into bedside utility, the researchers implemented the random forest algorithm within a user-friendly web application designed for real-time risk stratification. Rather than relying solely on traditional pre-extubation assessments that often yield suboptimal accuracy late in the care pathway, this platform leverages early emergency department resuscitation data. By facilitating the upstream identification of dynamic weaning failure risk, the tool provides intensivists with patient-specific risk predictions to guide stratified intensive care unit management and optimize personalized weaning strategies for mechanically ventilated patients with acute exacerbations of chronic obstructive pulmonary disease.
Subject of Research: Machine learning prediction of extubation failure in acute exacerbation of chronic obstructive pulmonary disease
Article Title: Early prediction of extubation failure in AECOPD patients using emergency department data: a multicenter interpretable machine learning study
Article References: Zhou, J., Chen, S., Xu, C., & Wen, X. (2026). Early prediction of extubation failure in AECOPD patients using emergency department data: a multicenter interpretable machine learning study. BMC Emergency Medicine. https://doi.org/10.1186/s12873-026-01839-7
Image Credits: AI Generated
DOI: 10.1186/s12873-026-01839-7
Keywords: chronic obstructive pulmonary disease, extubation failure, mechanical ventilation, emergency department, machine learning, random forest, SHAP interpretability, critical care, Early, prediction, extubation, failure
News Source: Teresa Odom. (November 7, 2026). Machine Learning Predicts Extubation Failure in Critical COPD Emergency Resuscitation. Scienmag.
