Performance evaluation of selected decision tree algorithms for COVID-19 diagnosis using routine clinical data

Background: The novel 2019 Coronavirus disease (COVID-19) poses a great threat to global public health and the economy. The earlier detection of COVID-19 is the key to its treatment and mitigating the transmission of the virus. Given that Machine Learning (ML) could be potentially useful in COVID-19 identification, we compared 7 decision tree (DT) algorithms to select the best clinical diagnostic model. Methods: A hospital-based retrospective dataset was used to train the selected DT algorithms. The performance of DT models was measured using performance criteria, such as accuracy, sensitivity, specificity, receiver operating characteristic (ROC), and precision-recall curves (PRC). Finally, the best decision model was obtained based on comparing the mentioned performance criteria. Results: Based on the Gini Index (GI) scoring model, 13 diagnostic criteria, including the lung lesion existence (GI= 0217), fever (GI= 0.205), history of contact with suspected people (GI= 0.188), O2 saturation rate in the blood (GI= 0.181), rhinorrhea (GI= 0.177), dyspnea (GI = 0.177), cough (GI = 0.159), history of taking the immunosuppressive drug (GI= 0.145), history of respiratory failure (ARDS) (GI= 0.141), lung lesion situation (GI= 0.133) and appearance (GI= 0.126), diarrhea (GI= 0.112), and nausea and vomiting (GI = 0.092) have been obtained as the most important criteria in diagnosing COVID-19. The results indicated that the J-48, with the accuracy= 0.85, F-Score= 0.85, ROC= 0.926, and PRC= 0.93, had the best performance for diagnosing COVID-19. Conclusion: According to the empirical results, it is promising to implement J-48 in health care settings to increase the accuracy and speed of COVID-19 diagnosis. © 2021. Iran University of Medical Sciences