Search

The coronavirus disease 2019 (COVID-19) has overwhelmed healthcare systems and continues to pose a significant threat worldwide. Predicting disease severity would enhance treatment provision and resource allocation. Although multiple studies were conducted to assess COVID-19's severity using machine learning (ML) models, few studies focus on patient medication history and comorbidities. In this study, ML algorithms were trained using a comprehensive dataset comprising medication history, comorbidities, and clinical findings. Patient data was gathered from King Fahad University Hospital (KFUH) in Saudi Arabia (IRB#: 2021-05-480). The dataset comprised 622 positive COVID-19 with 49 features. Three experiments were conducted to train four ML algorithms, including random forest (RF), gradient boosting machine (GMB), extreme gradient boost (XGBoost), and extra trees (ET). Findings revealed that GBM outperformed other models with 96.30% accuracy, 95.80% precision, 97.64% recall, and 96.69% F-score, with 23 features. Moreover, the permutation feature importance technique suggested that the five most influential features for forecasting disease severity were “CRP level”, “CO2 level”, “SrCr”, “Tocilizumab”, and “Age”. In addition, the shapley additive explanation (SHAP) recommended that the “D-Dimer level”, “CrCl”, and “Hypertension” were also influential. The development of an effective GBM model has the potential to aid medical specialists in the assessment of disease severity. While several models take into account patient presentation and laboratory findings, this study is unique in its scope, considering a far more comprehensive patient profile. The developed model was able to accurately predict features that have been clinically shown to correlate with disease severity. Of interest the model was able to identify a pattern of association between the use of certain medications such and disease severity. We report that the use of beta blockers may be associated with reduced severity, whereas the use of immune modulating drugs namely tocilizumab appeared to be associated with poor disease outcomes in this patient population.