Diagnostic Stewardship of Blood Cultures in the Pediatric ICU Using Machine Learning.

Objective: The medical community recently experienced a severe shortage of blood culture media bottles. Rates of blood stream infection (BSI) among critically ill children are low. We sought to design a machine learning (ML) model able to identify children at low risk for BSI to improve blood culture diagnostic stewardship.

Methods: We developed and validated an extreme gradient-boosting ML classifier using an existing dataset of retrospective pediatric intensive care unit (PICU) patients from a single institution. Data from children aged 3 months to 18 years who had a blood culture collected within 24 hours of PICU admission were included. The first 80% of patients (1/1/2011-12/21/2018) were used for model training and the last 20% (12/22/2018-12/25/2020) for testing (temporal validation). All 121 variables from the original dataset (vital sign, laboratory, and other clinical data) were included as predictors. Negative predictive value (NPV) was the primary evaluation metric.

Results: Of the 3121 blood cultures obtained during 2320 PICU admissions (2100 unique children), 205 (6.6%) were positive. Model NPV was 0.997 in the training set and identified 667/2321 (28.7%) of negative cultures. NPV was 0.993 in the test set and identified 151/595 (25.4%) of negative cultures. The number needed to harm was 151 (151 negative blood cultures could be avoided for each false negative prediction). Key predictors included central line presence, temperature rate of change, and mean platelet volume.

Conclusions: We trained and validated an ML model that accurately predicted >25% of subsequently negative blood cultures. If implemented prospectively, such models could help reduce unnecessary blood cultures among low-risk children.