A predictive model for hospitalization and survival to COVID-19 in a retrospective population-based study

The severe acute respiratory syndrome coronavirus (SARS-CoV-2) causing coronavirus disease 2019 (COVID-19) is highly transmissible and has been responsible for a pandemic associated with a high number of deaths. The clinical management of patients and the optimal use of resources are two important factors in reducing this mortality, especially in scenarios of high incidence. To this end, it is necessary to develop tools that allow early triage of patients with the minimal use of diagnostic tests and based on readily accessible data, such as electronic medical records. This work proposes the use of a machine learning model that allows the prediction of mortality and risk of hospitalization using simple demographic characteristics and comorbidities, using a COVID-19 dataset of 86867 patients. In addition, we developed a new method designed to deal with data imbalance problems. The model was able to predict with high accuracy (89-93%, ROC-AUC = 0.94) the patient’s final status (expired/discharged) and with medium accuracy the risk of hospitalization (71-73%, ROC-AUC = 0.75). These models were obtained by assembling and using easily obtainable clinical characteristics (2 demographic characteristics and 19 predictors of comorbidities). The most relevant features of these models were the following patient characteristics: age, sex, number of comorbidities, osteoarthritis, obesity, depression, and renal failure.