The Impact of Missing Data Handling Techniques on the Accuracy of COVID-19 Mortality Prediction

COVID-19, a novel disease, had introduced a new problem in the healthcare industry, where due to the lack of prior experience with the disease, healthcare professionals were unable to accurately estimate the Mortality Chance (MC) of a patient, thus challenging their ability to prioritize treatments and resources based on the urgency of a case. Enhancing the accuracy of Machine Learning (ML) models in predicting MC and using better predictions is a solution to the problem. The dataset used in this paper had missing values in 543444 data points (approximately 95.9% of the dataset), rendering most of the samples unusable unless treated using Missing Data Handling (MDH) techniques. Several MDH techniques were used in this paper to enhance the accuracy of ML models. These techniques were tested by creating an optimized ML Pipeline containing Dimensionality Reduction techniques, Oversampling methods, and various ML algorithms. The different MDH techniques were compared using Quantitative Analysis through Receiver Operating Characteristic (ROC) curves, Area Under ROC curve (AUROC) values, Accuracy scores and p-values obtained from Student’s T-Test and DeLong’s Test. This paper also compares the top two performing models using Shapley Additive Explanations (SHAP) to understand feature importance.