An Online Risk Calculator for Rapid Prediction of In-hospital Mortality from COVID-19 Infection

Abstract

Prediction of mortality from COVID-19 infection might help triage patients to hospitalization and intensive care. To estimate the risk of inpatient mortality, we analyzed the data of 13,190 adult patients in the New York City Health + Hospitals system admitted for COVID-19 infection from March 1 to June 30, 2020. They had a mean age 58 years, 40% were Latinx, 29% Black, 9% White and 22% of other races/ethnicities and 2,875 died. We used Machine learning (Gradient Boosted Decision Trees; XGBoost) to select predictors of inpatient mortality from demographics, vital signs and lab tests results from initial encounters. XGBoost identified O2 saturation, systolic and diastolic blood pressure, pulse rate, respiratory rate, age, and BUN with an Area Under the Receiver Operating Characteristics Curve = 94%. We applied CART to find cut-points in these variables, logistic regression to calculate odds-ratios for those categories, and assigned points to the categories to develop a score. A score = 0 indicates a 0.8% (95% confidence interval, 0.5 – 1.0%) risk of dying and ≥ 12 points indicates a 98% (97-99%) risk, and other scores have intermediate risks. We translated the models into an online calculator for the probability of mortality with 95% confidence intervals (as pictured):Abstract Figuredanielevanslab.shinyapps.io/COVID_mortality/