Prediction of central line-associated bloodstream infection: focus on time of insertion.

Objective: Central line-associated bloodstream infections (CLABSIs) result in morbidity and mortality among hospitalized patients. Hospital interventions to reduce the incidence of CLABSI are often broadly applied to all patients with central venous access. Identifying central lines at high risk for CLABSI at time of insertion will allow for a more focused delivery of preventative interventions.

Design: This was an observational cohort study conducted at three hospitals including all patients who received central venous access. CLABSIs were identified using an institutional CLABSI database maintained by the hospital epidemiology team. Logistic regression (LASSO) and machine learning (random forest, XGboost) techniques were applied for the prediction of CLABSI occurrence, adjusting for selected patent and insertion-level characteristics.

Results: A total of 40,008 central venous catheters were included, of which 409 (1.02%) were associated with CLABSI. The random forest and the XGBoost models had the highest discrimination (Area Under the Received Operating Curve [AUC] 0.79) followed by LASSO (0.73). High illness severity, receipt of total parenteral nutrition, receipt of hemodialysis, pre-insertion hospital length-of-stay, and low albumin levels were all predictive of CLABSI occurrence. Precision for all models was poor owing to a high false-positive rate.

Discussion: CLABSI can be predicted based upon patient and insertion level factors in the electronic health record. In this study, random forest and gradient-boosted models had the highest AUC. Prediction cut-offs for the identification of CLABSI can be adjusted based upon the acceptable rate of false-positives for a given CLABSI preventative intervention.