(711) Predicting Pediatric Extubation Failure Using Machine Learning

Introduction: Balancing the risks of extubation failure and continued mechanical ventilation is central to ventilator liberation in the pediatric intensive care unit (PICU). Sensitive predictive models do not exist for pediatric extubation failure. Our objective was to leverage a local quality improvement database and create a machine learning model to accurately predict pediatric extubation failure.

Methods: Using the local database, we evaluated all extubations (N=1076) in a single PICU from November 2018 through May 2022. The database contained descriptive variables of: age, race, sex, reason for intubation, and certain co-morbid conditions (genetic syndromes, neurologic disorders, respiratory disorders, surgical and medical airway conditions, chronic positive pressure requirement). Available clinical variables included hours of invasive ventilation, illness severity scores, spontaneous breathing trial outcome, peri-extubation steroid exposure, and peri-extubation Glasgow Coma Scores. Extubation failure was defined as reintubation within 48 hours for any reason other than a planned procedure. A logistic mixed model was created for the outcome of extubation failure. The model was trained on the first 844 extubations and validated on the remaining 232.

Results: Considering the entire cohort, the extubation failure rate was 10%. Peri-extubation steroids were negatively correlated with extubation failure (R: -0.60, p: 0.028, 95% CI -1.13 to -0.06) and sum of co-morbid risk factors was positively associated with failure (R: 0.244, p: 0.03, 95% CI 0.02 – 0.47). The model had an AUROC of 0.728 (95% CI 0.62-0.84).

Conclusions: This single-center, retrospective cohort study sought to build a machine learning model to accurately predict extubation failure. A logistic mixed effects model was built using accessible patient historical patient information and was tested on an independent cohort with reasonable sensitivity when validated with an independent dataset. The model did not perform better than clinical assessment, however, when considering the cohort's extubation failure rate.