A tailored machine learning approach for mortality prediction in severe COVID-19 treated with glucocorticoids

BACKGROUND The impact of severe COVID-19 pneumonia on healthcare systems highlighted the need for accurate predictions to improve patient outcomes. Despite the established efficacy of glucocorticoids (GCs), variable patient responses are observed, and the existing clinical scores are limited in predicting non-responders. We propose a machine learning (ML) based approach to predict mortality in COVID-19 pneumonia treated with GCs. METHODS This is an ML-driven retrospective analysis involving 825 patients. We leveraged XGBoost to select the most appropriate features from the initial 52, including clinical and laboratory data. Six different ML techniques were compared. Shapley additive explanation (SHAP) values were used to describe the influence of each feature on classification. Internal validation was performed. RESULTS Nine key predictors of death were identified: increasing C-reactive protein (CRP), decreasing arterial partial pressure of oxygen to fraction of inspired oxygen ratio (PaO2/FiO2), age, coronary artery disease, invasive mechanical ventilation, acute renal failure, chronic heart failure, PaO2/FiO2 earliest value, and body mass index. Random forest achieved the highest test area under the receiver operating characteristic curve at 0.938 (95% CI 0.903-0.969). SHAP values highlighted age and PaO2/FiO2 improvement as the most influential features; the latter showed a higher impact than CRP reduction over time. CONCLUSION The proposed ML algorithm effectively predicts the risk of hospital death in COVID-19 pneumonia patients undergoing GCs. This approach can be adapted to datasets measuring similar clinical variables.