Predicting Extubation Failure in Critically Ill Adults Receiving Invasive Mechanical Ventilation: A Secondary Analysis of the Wean Safe Cohort

Abstract RATIONALE Critically ill patients who fail an extubation attempt face a higher risk of poor in-hospital outcomes, including prolonged ICU stay and increased mortality. However, it remains unclear how best to estimate the risk of extubation failure in patients receiving invasive mechanical ventilation. We conducted a secondary analysis of the WEAN SAFE cohort (a multinational study that described weaning practices) to derive and internally validate different machine learning algorithms to predict extubation failure. METHODS WEAN SAFE enrolled eligible participants in 481 ICUs in 50 countries, who had received mechanical ventilation for 2 calendar days or longer. Our secondary analysis included adult patients who were deliberately extubated, excluding those with orders to withhold re-intubation. The main outcome of interest was the composite of re-intubation or death within 7 days of extubation. The cohort was randomly split in train and test samples (using a 2:1 ratio). We fitted classification algorithms using patients’ baseline demographics, comorbidities, reason and severity of ICU admission, physiologic and ventilation parameters closest to extubation time, and center-level features. We report discrimination performance (area under the receiver operating characteristic curve; AUROC), feature importance (a relative measure of each predictor's role), and calibration measures (e.g., Brier score) in the test sample for each model. RESULTS Of 5869 patients in the original WEAN SAFE cohort, 3584 met inclusion criteria. We excluded 286 patients with a pre-existing order (made before or on the day of extubation) to withhold reintubation. Of 3298 patients, 2199 and 1099 were randomly split in the train and test samples respectively. Overall, women comprised 38% of the train sample; 65% of patients presented at least one baseline comorbidity before ICU admission. ICU admissions were mostly medical (66%), followed by urgent or planned surgery (15% and 9% respectively). The composite outcome of re-intubation or death within 7 days of extubation occurred in 19% of patients. Tested models showed similar performance, with AUROC estimates between 0.63 and 0.65; Brier score estimates were between 0.15 and 0.16 (see Table). The variables with the highest predictive importance were 1) frailty, 2) age, 3) respiratory rate (RR) and pre-extubation PaO2:FiO2 ratio, 4) cardiac arrest, and 4) lower-income country. CONCLUSION Different supervised learning algorithms have adequate performance to predict extubation failure among critically ill adult patients receiving invasive mechanical ventilation who undergo a planned extubation attempt. The use of these models in clinical practice and the enrichment of future studies should be further explored.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.