Modelos predictivos de mortalidad basados en aprendizaje automático para pacientes con fiebre hemorrágica de Crimea-Congo

The aim of this study was to identify the risk factors associated with mortality in patients diagnosed with Crimean-Congo hemorrhagic fever (CCHF) through the application of machine learning models and to evaluate their predictive performance.

The study included patients with a definitive diagnosis who were admitted to the Department of Infectious Diseases and Clinical Microbiology of Tokat State Hospital between February 1, 2011 and October 1, 2022 with suspicion of CCHF. Five models, namely XGboost, Logistic regression, Random Forest, LightGBM, and Gradient Boosting Classifier, were constructed using machine learning algorithms to predict mortality in CCHF patients. The performance of these models was subsequently evaluated.

A total of 1,881 cases of suspected CCHF were admitted to the hospital, of which 891 were confirmed, resulting in a fatality rate of 3.3%. In the study, the receiver operating characteristic (ROC) analysis was performed to predict the risk of mortality in CCHF patients with the XGboost, logistic regression, and Gradient Boosting Classifier models. The area under the curve (AUC) results were 0.849, 0.919, and 0.853, respectively. In the evaluation of the relative importance of the features of these models, platelet count, neutrophil-to-lymphocyte ratio (NLR), and neutrophil count were identified as being among the top five.

Statistically significant predictive models were created using machine learning techniques, specifically XGboost, logistic regression, and Gradient Boosting Classifier. The results of our analysis suggest that platelet count, NLR and neutrophil count are the most effective parameters for predicting mortality.