Machine Learning Based Classification Models for COVID-19 Patients

The SARS-CoV-2 pandemic has pushed the National Health Service to extraordinary pressure, causing situations of imbalance between the request and availability of assistance. When the number of patients exceeds the available resources, doctors need to establish priorities among the patients to be treated. This paper describes novel data-driven optimization models to support doctors’ decisions to solve one of the main problems encountered during the first months of the COVID-19 pandemic: predict the mortality risk for COVID-19 in order to address the most appropriate therapeutic path. The models are trained using clinical data obtained at the access to the Emergency Department of 150 SARS-CoV-2 infected patients admitted to ASST-Valcamonica (Brescia, Italy), in March 2020. To handle the uncertainty in data, we formulate robust and distributionally robust optimization models and compare their performance with other 31 different classification models from the literature, including decision trees, discriminant analysis, support vector machines, logistic regression, nearest neighbors, and naive Bayes. Numerical results show that robust formulations allow to achieve higher levels of accuracy with respect to the corresponding deterministic ones. The best prediction results are obtained with an optimized decision tree model, allowing to identify the most important factors. The tool can be used after triage to more accurately assess the severity of a COVID-19 patient’s condition, allowing doctors to optimize patient accommodation by identifying those in need of intensive care and those instead of sub-intensive care.