COV-TabNet: Investigating the Influence of Underlying Disease on COVID-19 Patients

Numerous studies have demonstrated that coronavirus disease (COVID-19) infected people with underlying illnesses are more seriously unwell and possibly high-risk, with greater fatality rates, making it difficult for clinicians to detect potentially high-risk patients early. In this study, a new deep learning interpretable prediction model COV-TabNet based on the TabNet model is proposed, the data preprocessing process discards the missing value processing method that comes with the TabNet model and replaces it with the MICE missing value filling method that has a better effect, and the data balancing correction is proposed. The SMOTE-ENN module is added to optimize the model's parameters using a Grid search for model parameter optimization. The model aims to investigate the impact of underlying diseases on the prognosis and clinical outcomes of COVID-19 patients, as well as predict whether COVID-19 patients are admitted to the hospital for death and intensive care unit (ICU), to aid physicians in identifying high-risk patients. Our proposed method was compared to AdaBoost, K-Nearest Neighbor (KNN), Random Forest, SVM, Logistic Regression, and TabNet models for performance evaluation, and the results of our study stated that the AUC score for prediction of mortality reached 0.988, with the three important characteristics affecting mortality being hypertension, atrial fibrillation, and previous myocardial infarction. The AUC score for predicting ICU admission was 0.970, with hypertension, other cardiovascular disorders (excluding AF and infarction), and peripheral arterial disease being the three most important factors influencing ICU admission. COVID-19-infected people with simultaneous hypertension and cardiovascular disease have a worse prognosis and should be continuously watched, according to the results, which are useful for physicians.