A Machine Learning Model Predicts Future Infections in Elderly Patients in the Intensive Care Unit Who Carry Carbapenem-Resistant Enterobacteriaceae by Analyzing Easily Accessible Electronic Medical Records.

Background: The prevalence of carbapenem-resistant Enterobacteriaceae (CRE) presents a significant challenge in clinical anti-infective therapy among older adults in intensive care unit (ICU). Therefore, finding valid methods to rapidly identify patients with a high risk of CRE infection is essential. Creating a fully automated score based on a machine-learning algorithm may be a method to quickly predict the incidence of subsequent CRE infection in CRE Intestinal Carriers among older adults in the ICU. Methods: Older patients with positive CRE rectal swab screening were identified using electronic medical records from April 1, 2020, and April 31, 2024. Intestinal carriers who acquired CRE infections were compared to those who did not develop CRE infections. A least absolute shrinkage and selection operator (LASSO) was used to screen for essential features associated with CRE infection. Finally, three features (mechanical ventilation ≥96h, tigecycline exposure, SOFA score) were used to establish models. Four models, logistic regression model (LR), decision tree (DT), naive Bayes model (NBM), and support vector machine (SVM) classifier, were trained to establish a prediction model and a nomogram. The model's discriminatory capability was evaluated by determining the area under the curve (AUC). Additionally, calibration, decision curve analyses (DCA), and the bootstrapping validation method were conducted to assess the model further. Result: Of the 1433 patients, CRE colonized 71, and 28 developed CRE infection, of whom 13 developed CRE bloodstream infection (BSI). LR outperformed other classifier models in data and achieved the lowest Brier Score. AUC and the bootstrap value of the developed nomogram were 0.862(95% confidence interval [CI]: 0.780–0.943) and 0.862 (95% CI: 0.776–0.933), respectively. Besides, the calibration curve showed good consistency between the actual diagnosed CRE infection and the predicted probability, and DCA showed that if the threshold probability of CRE infection was between 42% and 98%, using the proposed nomogram to predict subsequent CRE infection in CRE intestinal carriers would obtain a net benefit. Conclusions: Our analysis revealed that this model would help in the individualized evaluation of the necessity of CRE de-colonization and inform strategies to eliminate the need for CRE infection.