694

Abstract

Introduction: Existing models to predict intensive care unit (ICU) mortality evaluate severity of illness using summarized data, which fails to integrate information related to evolving patient physiology. These models also often require input from a physician or a professional coder. Hypothesis: Machine learning advances can be leveraged to integrate variations in vital signs over the first 24 hours of admission to predict ICU mortality in a robust, fully-automated manner using data that are routinely collected, and without any burden on patients or caregivers. Methods: The census of patients admitted to the medical, surgical, cardiac and neurological ICU’s at a single, tertiary care center between 1/1/11 and 5/19/12 was evaluated retrospectively. The vital signs (systolic, diastolic and mean arterial pressure, heart rate, respiratory rate and oxygen saturation) from the first 24 hours were sampled each minute and uploaded into the electronic medical record. Demographic data (age, race, and gender) as well as Glasgow Coma Scale (GCS) were also recorded. Patients were stratified by percentage of recorded vital signs: >90%, >50%, all patients. An area under the receiver operating characteristic curve (AUC) maximizing support vector machine (SVM) model was trained using data collected before 12/14/11 to predict ICU mortality using summary features derived from the signals, demographic features, and GCS. The model was validated on data from patients admitted after 12/14/12. Results: The study cohort comprised 5,623 patients [mean age of 61 (+/- 15.7); 56% male]. The average ICU length of stay was 8.7 days. With all patients included the AUC was 0.794. This increased to 0.845 when over 50% of necessary data was available. Patients with > 90% of data available had an AUC of 0.850 with sensitivity of 0.876. Conclusions: Advances in machine learning provide the ability to assess patients in a fully-automated manner while factoring in changes in patient condition over the course of ICU admission. This approach can be used to predict ICU mortality robustly while leveraging data that is already available routinely, and without the need for coded values from human experts or physician intervention.

Ähnliche Arbeiten