Modelagem e validação de classificador para predição de sucesso no desmame da ventilação mecânica invasiva

Abstract

About 40% of severely ill patients require mechanical ventilatory support, and although invasive mechanical ventilation (IMV) is an essential therapy for patients with respiratory failure, it is associated with several complications. Thus, as soon as possible, patients should undergo ventilator weaning, a process that can last 40% of the total time required for mechanical ventilation. To consider successful weaning, the patient should maintain spontaneous ventilation for at least 48 hours after cessation of artificial ventilation. However, if return to ventilatory support is required within 48 hours after extubation, it is termed weaning failure. The Spontaneous Breathing Test (SBT) is still the most used technique to perform ventilatory weaning. However, the methods for performing this test did not present statistically significant differences for weaning success. Although recommended, it is important to note that, in studies, this test has not been so accurate and does not identify approximately 15% of weaning failures. The advent of the use of weaning indices may reduce the time spent on ventilation, however, the accuracy of these indices is still questionable, especially in individual patient profiles, which require the combination of clinical criteria and adequate parameters to conduct weaning properly. The aim of this study was to propose and validate the Classifier for predicting successful weaning from invasive mechanical ventilation in Intensive Care Unit patients. This is an observational, longitudinal, prospective, quantitative, exploratory and inferential study. A data collection form was used, divided into the following moments: 48 hours after the establishment of the IMV, at the moment before the completion of the SBT and after the removal of the IMV until the success or failure of weaning. The set of 214 patients was randomly divided into two groups: training group (n = 174) and test group (n = 40). With the data from the training group we used the k-fold cross-validation statistical method, with Firth penalty to find the best model for predicting weaning success and then for the Classifier, a cutoff point was established. using the Youden index, which summarized the probabilities returned by the logistic model equation in a binary prognosis “success” or “failure”. The variables that were statistically significant (p-value < 0.05) were: rapid shallow breathing index (RSBI) (OR= 0.82; AUC= 0.93, 95% CI 0.90 - 0.97) , arterial carbon dioxide partial pressure (PaCO2) (OR= 0.23; AUC= 0.90, 95% CI 0.86 - 0.95), ratio of oxygen partial pressure to inspired oxygen fraction (PaO2/FiO2 index) (OR= 0.23; AUC= 0.97, 95% CI 0.95 ‒ 1.00) and days of mechanical ventilation (OR= 0.38, AUC= 0.94, CI 95 % 0.90 - 0.97). The Classifier presented AUC= 1 and was tested with a new data set (test group) and presented accuracy= 0.95, Kappa= 0.94, sensitivity= 1.00, specificity= 0.92, positive predictive value = 0.96, negative predictive value= 1.00. The results of this study confirmed the hypothesis that the clinical variables of patients who were strongly associated with weaning success, when combined in a Classifier, constituted an index with greater predictive capacity for weaning success. Thus, this validated Classifier is useful in identifying patients who will succeed in weaning and its adoption in clinical practice is recommended to guide strategies and behaviors in predicting weaning success.