A. N. Kothari1, J. Attisha2, S. A. Brownlee1, A. Cobb1, C. Fairman1, K. Halvorsen1, W. Hopkinson1, H. H. Ton-That1, P. C. Kuo1 1Loyola University Chicago Stritch School Of Medicine,Surgery,Maywood, IL, USA 2DePaul University,College Of Computing And Digital Media,Chicago, IL, USA
Introduction:
Unplanned intubations are associated with significant patient morbidity and mortality. Given the important impact of unplanned intubations on both ventilator days and mortality, reductions in the number of unplanned intubations has far-reaching quality improvement implications. The overarching goal of this project was to create a mobile application for use during bedside rounds to identify patients at high daily risk for unplanned intubation.
Methods:
A single-center, retrospective review of patients who underwent a surgical procedure that met National Surgical Quality Improvement Program inclusion criteria and subsequently received care in the Surgical Intensive Care Unit (SICU) was conducted. The primary predicted outcome was an unplanned intubation. For each machine learning algorithm tested, a grid search was performed to identify the best hyper-parameter values. Each model was trained and tested 20 times on random balanced subsets of the dataset. The following performance metrics were used to choose the final model: accuracy, sensitivity, and specificity. Mobile application usability was measured using the System Usability Scale (SUS).
Results:
A total of 24,198 surgical encounters meeting NSQIP inclusion criteria were identified from January, 2012 – July, 2015. Only patients that received care in the SICU were included in the final analytic cohort (n=4,487). The overall unplanned intubation rate was 6.6%. A total of 8 different machine learning algorithms were tested: Decision Trees, Naïve Bayes, K-Nearest Neighbors, Linear Discriminant Analysis, Support Vector Machines, Random Forests, Stacked Ensemble Method, and Bucket Ensemble Method. Model performance of each is shown in Table 1. The Stacked Ensemble algorithm was integrated into a mobile application for bedside use and received a mean SUS score of 74 (SD 11.2, usability above average) by a pilot group of surgical faculty, residents, and intensive care nurses.
Conclusion:
The use of machine learning algorithms can create a high-performing predictive tool. Accurate prediction of high-risk individuals in the SICU using a mobile application will allow for the implementation of targeted interventions to better assist those patients most vulnerable to having unplanned intubation occurrences.
