Interpretable Machine Learning for Early Prediction of Multidrug-Resistant Organism Infection：A Discovery and Validation Study

Abstract

Abstract Background Multidrug-resistant organisms (MDRO) infection is a major public health threat in the world. We aim to predict risk of MDRO infections in Intensive Care Unit (ICU) patients by developing and validating a machine learning (ML) model.Methods This study included patients in the ICU from January 1, 2020 to December 31, 2022, and retrospectively analyzed the clinical characteristics of the patients. Lasso regression was used for feature selection. We use 6 machine learning methods to analyze clinical features and build prediction models. Furthermore, we illustrate the effects of the features attributed to the model and interpret the prediction process based on the SHapley Additive exPlanation(SHAP).Results A total of 888 cases were collected, 63 cases were excluded based on inclusion and exclusion criteria, and 825 final cases were included in the analysis, of which 375 were MDRO-infected patients. A total of 45 clinical variables were collected, and after selection, 31 variables were associated with outcomes and were used to develop machine learning models. We have build six ML models to predict MDRO infections, among which, the Random Forest (RF) model performs the best with an AUC of 0.83 and an accuracy of 0.767.Conclusions We built and validated an ML model for predicting patients who will develop MDRO infections, and the SHAP improves the interpretability of machine learning models and helps clinicians better understand the mechanisms behind the results. The model can provide guidance to ICU healthcare professionals in the prevention and control of patients at high risk of infection.