Determination of biochemical parameter dominance in the diagnosis of COVID-19 using machine learning methods

COVID-19 is a disease that appeared in 2020 and became a pandemic in a short time. Cough, shortness of breath, etc. - this disease, which appears with symptoms, has become a deadly disease as its effectiveness increases with time. For this reason, early detection of Covid-19 disease is important. Various laboratory tests, imaging techniques, etc. are used to diagnose Covid-19 disease. The real-time polymerase chain reaction test (RT-PCR) is considered the gold standard and is widely used. When diagnosing by laboratory tests, the workload increases due to the density of parameters, which makes diagnosis difficult in this case. Therefore, at the stage of diagnosis to specialists, an additional decision support mechanism is required. In this work, the system for assisting experts in decision making in the case of Covid-19 disease was developed using machine learning algorithms. Machine learning plays a supporting role for experts in many fields thanks to its wide range of applications. Some commonly used machine learning algorithms include the K-nearest Neighbor algorithm, support vector machines, decision trees, and artificial neural networks. To avoid problems arising from the redundancy of parameters, the number of parameters was reduced to an optimum by using feature selection methods in conjunction with machine learning algorithms. In this case, the specialists are relieved in the diagnosis phase with lower costs. The study analyzed the effectiveness of each parameter in diagnosing Covid-19 disease using 13 different feature selection methods for biochemical parameters. A total of 221 patient data consisting of 16 characteristics were used in the study, including 100 negative and 121 positive from Elazığ Fethi Sekin City Hospital. The number of parameters was reduced by using attribute selection methods and the classification process was performed. As a result of the study, it was found that the newly obtained feature set from the process performed with all features was successful. The 5 features obtained by feature selection methods gave the best performance for Artificial Neural Networks by feature selection method with Adaptive Structural Learning with an accuracy rate of 91.3%. It is expected that the study will help clinicians in clinical trials and provide information on the studies to be conducted. Keywords: Covid-19, Machine Learning, Feature Selection, Classification