Prediction of Decompensation in Patients in the Cardiac Ward

This study focuses on detecting deterioration of acutely ill patients in the cardiac ward at the University of Virginia Health System. Patients in the cardiac ward are expected to recover from a variety of cardiovascular procedures, but roughly 5% of patients deteriorate and have to be transferred to the Intensive Care Unit (ICU). Previous work has shown that early warning scores utilizing vitals signs and common lab results greatly lower morality for high risk patients. To build upon these results, data were collected over the course of two years from 71 beds in three cardiac-related wards at the University of Virginia Health System. In addition to information commonly collected for early warning scores, these data also contained continuous electrocardiography (ECG) telemetry data for all patients. Given that only one percent of observations are labeled as events, the F1 score was used as the primary metric to assess the performance of each model; area under the curve (AUC) was also considered. Previous work includes the development of logistic regression models with these data resulting in an AUC of 0.73. In this work, a super learner was built to further the study by stacking logistic regression, random forest, and gradient boosting models. Furthermore, a denoising auto-encoder was created to generate computer-derived features, the results of which were fed to machine learning models mentioned previously to predict patient deterioration. The logistic regression model built on existing and computer-generated features resulted in an F1 score of 0.1 and AUC of 0.7, which is comparable to previous models built on the same patient data set. The super learner had an improvement over existing logistic regression models, with an F1 score of 0.24 and AUC of 0.79.