Data-driven prediction of decannulation probability and timing in patients with severe acquired brain injury

Abstract

Background and objectives: From a rehabilitation perspective, removal of tracheostomy in patients with severe acquired brain injuries (sABI) is a crucial step. Predictive parameters for a successful decannulation are currently still a focus of the research for sABI patients, especially for those presenting a disorder of consciousness. For this reason, we adopted a data-driven approach predicting decannulation probability and timing using ensemble learning models in patients in intensive rehabilitation units.

Methods: 327 patients, 186 of which were successfully decannulated during their intensive rehabilitative stay, were recruited in a non-concurrent retrospective study. Decannulation probability and timing were predicted using data available within one week from admission at the rehabilitation unit. Two predictive models were trained and cross-validated independently, with the first being an ensemble of a Support Vector Machine and Random Forests and the second an Adaptive Boosting with a Support Vector Regression as weak learner. Confusion matrix, accuracy and AUC were considered as evaluation metrics for the classifier and median absolute error was considered for the regressor. To quantify the advantages in the clinical practice of using the latter prediction, we compared timing estimation with a timing guess (median) calculated on available data. The comparison was based on a Wilcoxon signed rank test.

Results: Decannulation probability was successfully predicted with an accuracy of 84.8% (AUC = 0.85) and timing with a median absolute error of 25.7 days [IQR = 25.6]. This resulted in a significant improvement with respect to the weaning time guess (p<0.05) with an effect size of 71.7%. Furthermore, dichotomizing the regression prediction with a threshold (3 months from the event), resulted in a prediction accuracy of 77.5% (AUC = 0.82) on the test set.

Discussions: A model capable of providing a prediction on decannulation probability and timing was developed and cross-validated, built on data taken at admission to the intensive rehabilitation unit. Translated in clinical practice, this information can support the clinical decision process and provide a mean to improve both in-hospital and domiciliary care organization.

Keywords: Decannulation; Disorder of consciousness; Machine learning; Prognostic models; Rehabilitation; Severe acquired brain injuries.