Work

Machine Learning–Based Prediction of Mortality and Multidrug-Resistant Infection in Non-Surgical ICU Patients with Suspected Infection: A Second Analysis of A Prospective National Multicenter Cohort Study

Year: 2026
Type: preprint
Abstract: Background Multidrug-resistant infection raise global disease burden. We aimed to develop and validate machine learning predictive models to identify independent risk factors for mortality and multidrug-resistant infection in non-surgical patients with suspected infection upon ICU admission in mainland China, using data from a prospective national multicenter cohort study. Methods This secondary analysis used prospectively collected medical data from non-surgical patients with suspected infection admitted to ICUs across mainland China between July 2021 and December 2022, derived from a national multicenter cohort study. Patients retained the parent study’s 7:3 random split into training and validation cohorts. Machine learning algorithms were applied to identify risk factors and develop predictive models for mortality and multidrug-resistant infection. Performance of logistic regression, random forest, extreme gradient boosting, and gradient boosting machine models was assessed using AUC, sensitivity, specificity, PPV, NPV, and Youden index. Results A total of 2270 patients from 67 hospitals were enrolled. Pulmonary (83.74%) were predominant, with an overall mortality of 9.74%. The logistic regression achieved an AUC of 0.93 for mortality prediction, identifying APACHE Ⅱ score( p < 0.01), and K. pneumoniae infections( p = 0.03) as independent risk factors. It also yielded the optimal AUC of 0.86 for multidrug-resistant infection prediction. Nine independent risk factors were confirmed, including: bloodstream infections( p = 0.01), urinary infections( p = 0.03), central system infections( p = 0.02), K. pneumoniae infections( p < 0.01), Stenotrophomonas maltophilia infections( p < 0.01), Escherichia coli infections( p < 0.01), Pseudomonas aeruginosa infections( p < 0.01), Enterococcus faecium infections( p < 0.01), and Staphylococcus epidermidis infections( p < 0.01) for non-surgical ICU patients with suspected infections. Conclusions Among non-surgical ICU patients with suspected infection, the logistic regression model identified subgroup-specific predictors for mortality and MDR infection, which differed from the overall cohort. This model may assist risk stratification and antimicrobial stewardship in medical critically ill patients.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7225915066
Cites: 35
Cited by:
Related to: 10