LAUSR

Introduction/Hypothesis: Conventional weaning protocols using a spontaneous breathing trial are often used for extubation and liberation from ventilatory support among patients receiving mechanical ventilation. However, after a successful SBT and extubation, 10-25% of patients require reintubation. In this study, we aimed to apply machine learning methods predict successful extubation. Methods: A retrospective observational study was conducted in 2018. Adult patients who underwent mechanical ventilation for at least 12 hours were enrolled in this study. A machine learning method using random forest model was developed to predict successful extubation. Demographics, vital signs, laboratory data, and data associated with mechanical ventilator were extracted from medical records, and we labelled the data for 2 hours after intubation as impossible extubation and the one for 3 hours before extubation as possible extubation. The dataset was randomly divided into a derivation set(70%) used to train the model and a validation set(30%) used to test the accuracy of the model. In the derivation dataset, the random forest model was used to predict the probability of successful extubation. Subsequently the area under the receiver operating characteristic curve(AUC), sensitivity, and specificity for the outcome in the validation dataset were obtained to predict successful extubation. Results: Of the 39 patients included in this study(median age, 67 [IQR:46-79] years; 25[64.1%] male patients), 35 patients were successfully extubated, and 4 patients required reintubation after extubation. Median length of mechanical ventilation in the successful extubation group and in the reintubation group were 4.0 days and 4.5 days, respectively. The AUC in the random forest model to predict successful extubation were 0.89 (95% confidence interval[CI]:0.84-0.95). The sensitivity and specificity values were 87.4% and 71.0%, respectively. Maximum airway pressure, mean airway pressure, PEEP were the important predictors of successful extubation, and the relative values based on 1 as the maximum value of the variable importance were 1.00, 0.86, 0.70, respectively. Conclusions: The machine learning for prediction successful extubation shows significantly high AUC scores, so that it can be applied to the assessment of extubation in clinical practice.