Background
Invasive candidiasis is a common infectious complication in geriatric critical care patients, which is thought to be related to malnutrition. The geriatric nutritional risk index (GNRI) score is an integrative and convenient tool to dynamically and comprehensively assess a patient’s current nutritional status. Thus, this study aims to assess the association between GNRI scores and invasive candidiasis in geriatric critical care patients.
Methods
A total of 5390 patients from the MIMIC-IV database were included in the training cohort to develop disease prediction models by logistic regression, Gradient Boosting Machine (GBM), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Random Forest (Rf) model. Developed models were validated in the test cohort consisting of 2766 patients from the MIMIC-III database. The area under the receiver operating curve (AUC) is used to represent the predictive performance of the different models.
Results
A decreased GNRI was significantly associated with invasive candidiasis in geriatric critical care patients (P < 0.001). The AUCs for the Rf model and GBM model were not significantly different than that of the logistic regression model (0.7093 versus 0.683; P = 0.4562; 0.6874 versus 0.683; P = 0.9178) in test cohort. The AUCs for the XGBoost model and LightGBM model were significantly lower than that of the logistic regression model (0.511 versus 0.683; P < 0.001; 0.6874 versus 0.592; P < 0.001) in the test cohort. Further analysis showed that GNRI as a continuous variable rather than a categorical variable is more valuable for predicting invasive candidiasis in our cohort.
Conclusion
Lower GNRI score was significantly associated with an increased risk of invasive candidiasis in geriatric ICU patients. Machine learning, particularly the GBM model and Rf model, can help in the prediction of invasive candidiasis.