Predicting mortality in intensive care unit patients infected with Klebsiella pneumoniae: A retrospective cohort study

Introduction: Although several models to predict intensive care unit (ICU) mortality are available, their perfor-mance decreases in certain subpopulations because specific factors are not included. Moreover, these models often involve complex techniques and are not applicable in low-resource settings. We developed a prediction model and simplified risk score to predict 14-day mortality in ICU patients infected with Klebsiella pneumoniae. Methodology: A retrospective cohort study was conducted using data of ICU patients infected with Klebsiella pneumoniae at the largest tertiary hospital in Northern Vietnam during 2016-2018. Logistic regression was used to develop our prediction model. Model performance was assessed by calibration (area under the receiver operating characteristic curve-AUC) and discrimination (Hosmer-Lemeshow goodness-of-fit test). A simplified risk score was also constructed. Results: Two hundred forty-nine patients were included, with an overall 14-day mortality of 28.9%. The final prediction model comprised six predictors: age, referral route, SOFA score, central venous catheter, intracerebral haemorrhage surgery and absence of adjunctive therapy. The model showed high predictive accuracy (AUC = 0.83; p-value Hosmer-Lemeshow test = 0.92). The risk score has a range of 0-12 corresponding to mortality risk 0-100%, which produced similar predictive performance as the original model. Conclusions: The developed prediction model and risk score provide an objective quantitative estimation of individual 14-day mortality in ICU patients infected with Klebsiella pneumoniae. The tool is highly applicable in practice to help facilitate patient stratification and management, evaluation of further interventions and allo-cation of resources and care, especially in low-resource settings where electronic systems to support complex models are missing.