? Machine learning application to predict the diagnosis of covid19 based on symptomatological patterns

ALBORNOZ, GERMÁN; MERCEDES DIDIER GARNHAM; MARCELA PILLOFF; MARINA VALERIA MOZGOVOJ; DUS SANTOS MARIA JOSE

Congreso; XI Argentine Congress of Bioinformatics and Computational Biology (XI CAB2C); 2021

Background:A new type of coronavirus was identified in 2019 in the city of Wuhan, China. This new virus was called SARS-CoV-2 and is the etiological agent of the acute respiratory disease known as COVID-19, which spreaded around the world causing a pandemic. This disease affects patients differently, some present symptoms similar to a regular flu, such as fever, cough, odynophagia, myalgia, headache, among others. Other patients present severe symptoms, like bilateral pneumonia, severe acute respiratory syndrome, and septic shock. One of the research fields that currently stands out in the fight against COVID-19 is Artificial Intelligence. Through the usage of computational models capable of learning to recognize patterns in a set of data, the symptomatic patterns of patients can be analyzed in order to determine which set of symptoms are representative of positive COVID-19 cases and thus, could be used to predict the diagnosis. From a clinical and epidemiological surveillance point of view, it is very important to know what specific symptoms are associated with the disease. In this work, we selected 1500 symptomatic cases from health centers located in the districts of Hurlingham and Ituzaingó. Nasopharyngeal swabs from these cases had been processed at the COVID Unit belonging to the National University of Hurlingham (UNAHUR). A dataset with the symptomatic pattern and their result for SARS-CoV-2 using RT-qPCR was created (n=750 positive patients and n=750 negative for COVID-19) and a Random Forest classification algorithm was used to analyze them. Results:The Random Forest analysis showed a 77% effectiveness and a 73% accuracy to discriminate between positive and negative cases. The sensitivity and the precision for the detection of positive cases was 93% and 70%, while for negative cases the values were 59% and 90% respectively. According to the model, the most relevant synthem to discriminate was the fever.Conclusions:The application of Machine Learning models, such as Random Forest, allows an accurate selection of positive cases and the prioritization of the patients to be tested. This would contribute with an efficient diagnostic procedure and the consequent optimization of available resources.Keywords: COVID-19, Machine Learning, Symptomatic patterns