Outcome Prediction of Covid-19 Patients from Clinical Admission Data Using Machine Learning Models

The coronavirus pandemic created a collapse in most healthsystems among other causes, due to the low availability of intensive careunits (ICU). As a consequence, some patients ended in a fatal outcome. Some initial studies showed that the presence of underlying conditionscould be predictors of the complexity of the disease. However, this isfar from conclusive. This article proposes the use of strategies based onarticial intelligence to predict the outcome of the disease in patientsdiagnosed with COVID-19 from data and results of para-clinical testsobtained at the time of hospital admission. The main goal for the modelis developing the ability to identify patients who are going to presentcomplications (admission to the ICU or death), from those who are dischargedwithout having gone through this unit. For this, four characteristicsselection techniques are evaluated, which allow establishing thesubset of characteristics that provide the most information to the decision,and three prediction models based on supervised classicationtechniques: Random Forest, support vector machines (SVM) and multikernellearning (MKL). The assessment was performed on a subset ofdata from HM-hospitals' COVID Data Save Lives initiative. The highestsensitivity was reported by the combination of a subset of the characteristicsselected with the mutual information method, and an SVMclassication model; while the highest specicity was obtained by theMKL model (87 %) with a set of characteristics selected by this samemodel.