Predicting COVID-19 mortality in Zambia - an Application of Machine Learning

Background:The Corona virus, has caused havoc all over the world, it has left no country untouched resulting in millions of cases and deaths. In an effort to fight back, scientist and public health professionals have used every form of advancing technology to curb the spread, predict the unforeseen adverse events, improve preparedness, and bring the world under control once more.Objective:The objective of this study was to predict mortality in hospitalized COVID-19 patients in Zambia using ML methods from a number of predictors that have been shown to be predictive of mortality.Methods:This research used powerful ML models in predicting COVID-19 mortality in 1,433 hospitalized patients in Zambia. The feature importance analysis helped in identification of important factors. The ML models GB, RF, SVM, DT, LR, and NB were used the performance metrics checked for each model were accuracy, recall, specificity, precision, F1 Score, ROC-AUC, and PRC-AUC.Results:The feature importance analysis found that hospital length of stay (LOS) and white blood cell count were the most influential features, other factors arranged in order of reducing importance included: age, wave, diabetes, hypertension, and sex. The GB achieved accuracy of 91.5%, recall of 93.6%, F1 Score of 91.7%, and ROC-AUC of 96.9%. The RF achieved accuracy of 90.9%, recall of 93.8%, F1 Score of 91.2%, and ROC-AUC of 96.8%. The SVM achieved accuracy of 87.8%, recall of 91.2%, F1 Score of 88.2%, and ROC-AUC of 94.1%. The accuracy and ROC-AUC of other models were 88.2% and 90.7% respectively for DT, 81.9% and 90.1% respectively for LR, and 79.2% and 86.9% respectively for NB.Conclusion:The study successfully derived and validated multiple ML models that predicted mortality effectively with reasonably high performance in stated metrics. The GB was the best suited for the data in our study. GB was thus recommended for similar studies with RF as best alternative. Knowledge of underlying health conditions about patients (length of hospitalization (LOS), white blood cell count, age, sex, hypertension, diabetes, and other factors) can help healthcare providers offer lifesaving services on time, improve preparedness and decongest health facilities.