Machine Learning Models to Establish the Risk of Being a Carrier of Multidrug-Resistant Bacteria upon Admission to the ICU.

Abstract

Objectives: To establish the risk of being a carrier of multidrug-resistant bacteria (MDR) upon ICU admission, according to the risk factors (RFs) from the Spanish "Resistencia Zero" (RZ) project checklist, using machine learning methodology. Methods: A retrospective cohort study, conducted with a consecutive sample of patients admitted to the ICU between 2014 and 2016. The study analyzed the RZ RFs for MDR, as well as other pathological variables and comorbidities. The study group was randomly divided into a development group (70%) and a validation group (30%). Several machine learning models were used: binary logistic regression, CHAID-type decision tree, and the XGBOOST methodology (version 2.1.0) with SHAP analysis. Results: Data from 2459 patients were analyzed, of whom 210 (8.2%) were carriers of MDR. The risk grew with the accumulation of RF. Binary logistic regression identified colonization or previous infection by MDR, prior antibiotic treatment, living in a nursing home, recent hospitalization, and renal failure as the most influential factors. The CHAID tree detected MDR in 56% of patients with previous colonization or infection, a figure that increased to almost 74% if they had also received antibiotic therapy. The XGBOOST model determined that variables related to antibiotic treatment were the most important. Conclusions: The RZ RFs have limitations in predicting MDR upon ICU admission, and machine learning models offer certain advantages. Not all RFs have the same importance, but their accumulation increases the risk. There is a group of patients with no identifiable RFs, which complicates decisions on preventive isolation.