Application of machine learning tools for feature selection in the identification of prognostic markers in COVID-19

Abstract Objective To identify prognostic markers by applying machine learning strategies to the feature selection. Methods An observational, retrospective, multi-center study that included hospitalized patients with a confirmed diagnosis of COVID-19 in three hospitals in Colombia. Eight strategies were applied to select prognostic-related characteristics. Eight logistic regression models were built from each set of variables and the predictive ability of the outcome was evaluated. The primary endpoint was transfer to intensive care or in-hospital death. Results The database consisted of 969 patients of which 486 had complete data. The main outcome occurred in 169 cases. The development database included 220 patients, 137 (62.3%) were men with a median age of 58.2, 39 (17.7%) were diabetic, 62 (28.2%) had high blood pressure, and 32 (14.5%) had chronic lung disease. Thirty-three variables were identified as prognostic markers, and those selected most frequently were: LDH, PaO2/FIO2 ratio, CRP, age, neutrophil and lymphocyte counts, respiratory rate, oxygen saturation, ferritin, and HCO3. The eight logistic regressions developed were validated on 266 patients in whom similar results were reached (accuracy: 65.8–72.9%). Conclusions The combined use of strategies for selecting characteristics through machine learning techniques makes it possible to identify a broad set of prognostic markers in patients hospitalized for COVID-19 for death or hospitalization in intensive care.