Development of a predictive model for Extubation Failure in weaning from mechanical ventilation: A retrospective cohort study

Extubation failure (EF) occurs in 2–25% of ICU patients. Our objective was to develop an EF predictive model. We performed a retrospective cohort study in a medical-surgical ICU with 40 beds at a University Hospital. Were analyzed 1017 patients, from January 2010 to December 2014, all over 16 years old, undergoing invasive ventilation for more than 24 h, and successful spontaneous breathing test (SBT). Seventeen variables were evaluated; we utilized logistic regression analysis with an evaluation of discrimination and calibration based on the area under the ROC curve (AUC-ROC) and the Hosmer-Lemeshow's goodness-of-fit test (Chi2 H-L), respectively. Extubation failure was present in 157 patients (15.4%); we developed a predictive model that included PaO2/FIO2 ratio ≤ 237.5, hemoglobin ≤9.5 g, accumulated fluid balance > 6022 ml, APACHE II > 16, blood urea nitrogen > 22.5 mg/dl and the presence of cardiopulmonary diagnostics. This model exhibited an AUC-ROC = 0.689 and a Chi2 H-L, p = 0.579. This study presents a risk score with an estimated probability of EF based on a multivariate predictive model. Due to the strong limitation of our retrospective study, however, it is necessary for an independent prospective cohort to improve discrimination and to prove the model applicability.