چكيده به لاتين
In recent years, the COVID-19 pandemic has placed many challenges on countries' health systems and treatment sectors. Despite being controlled, this disease caused a painful experience for people and the healthcare system worldwide and left behind the fear of similar epidemics of other emerging diseases. As a result, the countries' healthcare systems always try to have appropriate and efficient solutions to deal with these epidemics that can reduce the death rate of these diseases. This study aimed to develop a model to predict and identify patients with severe COVID-19 who are at risk of death. The model makes it easier for medical staff to make more accurate decisions and precisely prioritize patients in the challenging conditions of the epidemic. This research used retrospectively collected data from Shohada Tajrish hospital in Tehran, Iran. Forty-two demographic, clinical parameters, vital signs, blood test indices, and blood gas test indices were collected from COVID-19 patients admitted to the hospital from 22 July 2020 to 19 January 2021 to predict patients at risk of death by applying statistical and machine learning methods. Logistic regression, multilayer perceptron, support vector machine, and ensemble voting classifier outperformed other methods with 78% accuracy, recall 0.67, AUC 0.84, and F1-Score 0.70. Then the parameter set which was more related to the outcome was selected using a sequential forward feature selection algorithm. Correspondingly the best parameter set was chosen, which includes 13 parameters, including age, oxygen saturation measured by a pulse oximeter, albumin, blood urea nitrogen, white blood cells count, platelet count, lactate dehydrogenase, respiration rate per minute, oxygen saturation in venous blood gases test, prothrombin time, the base excess amount, creatinine and pH. Eventually, a reduced model was fitted using these 13 parameters. Logistic regression, AdaBoost, linear discriminant analysis, and ensemble voting classifier consisted of these three methods propose the best prediction with 81% accuracy, recall 0.77, AUC 0.84, and F1-Score 0.75.
