COST-SENSITIVE HETEROGENEOUS STACKED ENSEMBLE CLASSIFIER MODEL FOR PREDICTING SARS-COV-2 INFECTION FROM CLINICAL AND LABORATORY DIAGNOSTIC RESULTS

Abstract

The novel coronavirus disease (COVID-19) was initially identified in Wuhan, China, in December 2019 and subsequently had a profound global impact due to its rapid spread. Earlier symptoms and conditions of this deadly virus share common characteristics with the common cold and influenza, making the diagnosis difficult for health care professionals. This study aims to provide a decision on earlier and reliable screening of COVID-19 patients from the clinical blood tests and laboratory reports. We have proposed a cost-sensitive heterogeneous stacked ensemble classifier (CS-HSEC) that combines three different classifiers, namely random forest (RF), gradient boosting machine (GBM) and K-nearest neighbour (K-NN) at level-1 and naive bayes classifier (NB) as the meta classifier at level-2 for diagnosing COVID-19 disease. We have selected highly significant features using the hybrid feature selection algorithms based on mutual information (MI) technique to enhance the performance of the classifiers. The classifiers are trained and tested on the dataset acquired from the Israelita Albert Einstein Hospital, Sao Paulo, Brazil. The proposed CS-HSEC performance is evaluated using 10 clinical features of complete blood test with stratified 10-fold cross-validation and stratified train-test split, with 90% data for training and remaining for testing. The experimental results prove the proposed model's promising performance, having an accuracy of 98.33% with sensitivity, specificity and AUC of 0.93, 0.99 and 0.96, respectively. The proposed CS-HSEC can potentially find the COVID-19 disease during the initial screening of a patient and can be used as an assisting tool by clinicians in diagnosing the COVID-19 disease.