Mortality prediction with machine learning in COVID-19 patients

• Published in: Journal of critical care 2024-06, Vol.81, p.154589, Article 154589
• Main Authors: Yildirim, Suleyman, Sunecli, Onur, Kirakli, Cenk
• Format: Article
• Language: English
• Subjects: Coronaviruses; COVID-19; Disease severity score; Intensive care; Machine learning; Mortality
• ISSN: 0883-9441; 1557-8615
• DOI: 10.1016/j.jcrc.2024.154589

Mortality prediction is important for intensivists and conventional disease severity scores may fail to predict mortality in patients with coronavirus disease-19 (COVID-19) consistently. We aimed to develop a model using machine learning technology to predict mortality in patients with COVID-19 admitted to the intensive care unit (ICU). A total of 436 patients with COVID-19 who were followed up in the ICU between March 15, 2020, and December 31, 2021, were screened retrospectively. The worst laboratory results and vital signs of the patients were recorded in the first 24 h in the ICU. We selected a total of 29 inputs for develop a model by machine learning. A total 108 patients who were followed up between January 1, 2022, and March 31, 2022, were used for the testing model prospectively. Our model predicted mortality with 88% sensitivity and 88% specificity. Conventional disease severity scores predicted mortality with lower sensitivity and specificity than our model (71% sensitivity and 70% specificity for APACHE-2, 75% sensitivity and 75% specificity for SAPS-2 and APACHE−4). Our model had better discriminative power for mortality with area under curve, (AUC) 0.93 (95%CI, 0.87–0.98) than conventional disease severity scores (Fig. 1). Respiratory support on the first day of ICU admission was found to be the most important factors affecting mortality. Our model can predict ICU mortality with higher predictive power than the three conventional disease severity scores in patients with COVID-19. Conventional disease severity scores, such as APACHE-2 and SAPS-2, have been used for many decades for mortality prediction in ICU. Although conventional disease severity scores have good discriminative power for mortality, they seem to underestimate mortality (1). This situation may stem from the unique clinical course of COVID-19, increased patient load in intensive care units, or a lack of equipment and trained personnel resources in the ICU. In cases such as epidemics, where conventional disease scores are insufficient to predict mortality, models that can reliably predict disease outcomes can be developed with machine learning.