Abstract 17400: Using Electronic Health Records for Predicting Hospitalization of COVID-19 Patients in Massachusetts

Abstract

Introduction: The electronic health record (EHR) provides a platform to design models for predicting the severity of illness in COVID-19 patients. We used the Mass General Brigham (MGB) EHR that contains records mostly from greater Boston, Massachusetts. The objective of this research is to predict risk of hospitalization in patients who tested positive for COVID-19. Methods: Between 01/01/2020 to 05/11/2020 the MGB EHR contained 8,213 laboratory- confirmed COVID-19 patients. We queried the EHR for medical records of these patients from 12/31/2015 to 05/11/2020. To examine pre-COVID-19 medical history, we considered latest records prior to 21 days before the first COVID-19 positive test. To consider patients who were hospitalized before their test results become available, we analyzed admission records from 7 days before the first positive test. We used patients with no missing medical records to construct our dataset (N=2,783) and conducted a forward feature selection on 51 variables to identify markers that minimized the Akaike information criterion for predicting hospitalization. These markers then were used to train a generalized linear model on a subset of the dataset of MGB non-employees (N=2,367; 56% hospitalized). The trained predictive model was then tested on the remaining subset of the dataset that contains MGB employees (N=416; 17% hospitalized). Results: The model based on EHR data predicted risk of hospitalization well, with area under the curve (AUC, DeLong method) of 0.76 [95% CI: 0.7002-0.8206], Figure. Significant determinants of hospitalization risk included demographic variables (age, sex, race), prior medication use (heparin, erythropoietin, and beta blockers), and lab tests (albumin and white blood cell counts). Conclusions: EHR data of COVID-19 patients can be leveraged to predict risk of hospitalization among COVID-19 patients. Further validation of the predictive model on other cohorts is needed.