Determination of predictors of an unfavorable outcome in the subacute period of SARS-CoV-2 infection using machine learning methods
Автор: Dolgalev I.V., Vrazhnov D.A., Tolmachev I.V., Starikova E.G., Kaverina I.S., Zavyalova M.V.
Журнал: Сибирский журнал клинической и экспериментальной медицины @cardiotomsk
Рубрика: Цифровые технологии в медицине и здравоохранении
Статья в выпуске: 1 т.40, 2025 года.
Бесплатный доступ
Background. Pathological changes in systems and organs after COVID-19 can lead to delayed death. One of the most influenced target systems of post-COVID changes is the cardiovascular system.Aim: To identify, using machine learning methods, indicators that have predictive value in determining the adverse outcome of subacute COVID-19.Material and Methods. The study included 212 people admitted after previous severe COVID-19. Retrospectively, the patients were divided into 2 groups: 140 patients discharged from the hospital after improvement in their state and 72 patients died during hospitalization. All patients underwent general clinical, biochemical analyses, assessment of blood coagulation system. The following machine learning methods were used for data analysis: support vector machine, random forest, stochastic gradient boosting. Validation of the obtained models was carried out by the method of 10-fold cross-validation in conjunction with ROC-AUC analysis (Receiver Operation Characteristics - Area Under Curve).Results. In the created models, the predictors of mortality were urea and body temperature for the random forest and stochastic gradient boosting methods, erythrocyte, eosinophil and monocyte counts, and INR (International Normalized Ratio) level for the support vector machine.Conclusion. In our study, two predictive models created using machine learning methods random forest and stochastic gradient boosting showed that changes in urea level and body temperature had predictive value. The support vector machine revealed other predictors, namely the number of erythrocytes, eosinophils and monocytes, INR. We used the voting method, on the basis of which the urea level and body temperature were established as informative signs. The random forest and stochastic gradient boosting methods showed similar results, we did not take into account the data obtained using the support vector machine. This approach of choosing a predictive model by voting is often used when evaluating data using artificial intelligence methods. It is possible that an increase in urea levels was a trigger leading to endotheliitis and subsequent myocardial infarction, before acute renal failure developed.
Covid-19, post-covid changes, machine learning, urea, body temperature, erythrocyte count, eosinophil count, monocyte count, international normalized ratio, cardiovascular system
Короткий адрес: https://sciup.org/149147883
IDR: 149147883 | DOI: 10.29001/2073-8552-2025-40-1-199-208