Machine Learning-Based Prediction of Mortality and Multidrug-Resistant Infection Risks in ICU Patients with Suspected Infection: A Prospective National Multicenter Cohort StudyAuthor information

Background Suspected infection or infection may develop into sepsis or septic shock, leading to high mortality rate of patients admitted to ICU. However, suspected infection has not been fully characterized. We performed prediction models to identify independent risk factors of mortality and multidrug resistance for patients with suspected infection when they admitted to the ICU in mainland China. Methods A prospective analysis of Demographic, physiological and microbiological data were recorded for patients with suspected infection when they admitted to ICU between July 2021 and December 2022 in mainland China. Machine learning algorithms were employed to identify risk factors and create prediction models for mortality and multidrug resistance for patients with suspected infection. AUC were calculated and compared by bootstrap to evaluate prediction models. Results A total of 2963 patients from 67 hospitals in mainland China were enrolled into this study. The most common sites of infection were the lung (79.28%), bloodstream (17.11%) and abdomen (16.54%). The mortality rate was 10.90%. Logistic regression prediction model with AUC value 0.87 was selected and identified surgery ( p  < 0.01), APACHE Ⅱ ( p  < 0.01), and bloodstream infection ( p  < 0.01) were independent risk factors of mortality. Furthermore, logistic regression prediction model exhibited the highest AUC (0.86) for predicting the risk of multidrug-resistant infections and identifying six independent risk factors including APACHE Ⅱ ( p  < 0.01), bloodstream infections ( p  < 0.01), urinary infections ( p  < 0.01), Klebsiella pneumoniae ( p  < 0.01), Acinetobacter baumanii ( p  < 0.01), and Enterococcus faecium ( p  < 0.01). Conclusions The most common infection was pneumonia for patients admitted to ICU in mainland China. By means of machine learning techniques, we selected independent risk factors, as well as evaluated prediction models for the mortality and multidrug resistance of patients with suspected infection when they admitted to ICU.