Calculating the Risk of Admission to Intensive Care Units in COVID-19 Patients Using Machine Learning

Background: The COVID-19 pandemic clearly posed a global challenge to healthcare systems, where the allocation of limited resources had important logistical and ethical implications. Detecting and prioritizing the population at risk of intensive care unit (ICU) admission is the first step to being able to care for the most vulnerable people and avoid unnecessary consumption of resources by mildly ill patients. Objective: To create a model, using machine learning techniques, capable of identifying the risk of admission to the ICU throughout the hospital stay of the COVID patient and to evaluate the performance of the model. Methods: A retrospective cohort design was used to develop and validate a classification model of adult COVID-19 patients with or without risk of ICU admission. Data from three hospitals in Spain were used to develop the model (n = 1272) and for subsequent external validation (n = 550). Sensitivity, specificity, positive and negative predictive value, accuracy, F1 score, Youden index and area under the curve of the model were evaluated. Results: The LightGBM model, incorporating 40 variables, was used. The area under the curve obtained by the model when the test dataset was used was 1.00 (0.99-1.0), specificity 0.99 (0.97-1.00) and sensitivity 0.92 (0.86-0.98). Conclusions: A model for predicting ICU admission of hospitalized COVID-19 patients was created with very good results. The identification and prioritization of COVID-19 patients at risk of ICU admission allows the right care to be provided to those who are most in need when the healthcare system is under pressure.