Computational and statistical insights for leveraging biomarker-driven machine learning models in predicting COVID-19...

Global healthcare systems faced substantial challenges during the COVID-19 pandemic, which emphasized the requirement for precise methods to predict patient outcomes/mortality. The present work uses patients' biomarker data to evaluate the potential of 14 different machine learning (ML) models in predicting COVID-19 patient outcomes, i.e., survival vs. non-survival.

ML models were initially supplied with 26 clinical biomarkers of patients available in public repositories. Models' performance parameters such as accuracy, precision, recall, specificity, AUPRC, and AUC-ROC are evaluated, and then models are accordingly ranked to identify the five best-performing models. SHAP plots identified 15 common and most influential biomarkers (including two demographic features i.e., age and gender) regulating the patient outcome. Lower-ranked models and less-influential biomarkers were ablated. Five best-performing models, LightGBM, Random Forest, CatBoost, XGBoost, and Gradient Boost, are trained and tested using 15 most influential biomarkers on two patient cohorts separately.

Bland-Altman analysis affirmed that these models closely predict the clinical outcomes, i.e., survival and non-survival in patient groups of different ages and genders. In addition, partial dependence plots and paired t-tests displayed that ML models successfully captured biomarkers' clinical role in deciding patient mortality. LightGBM exhibited better performance than the other four ML models. Thus, based on day-wise biomarker data, identified ML models can be potentially utilized in predicting COVID-19 patient outcomes.

Overall, the outcomes can be beneficial in monitoring the disease progression, rapid resource allocation in high-mortality areas, and clinical management of critical patients.