Monitoring for early prediction of gram-negative bacteremia using machine learning and hematological data in the...

This study aims to develop an artificial intelligence-assisted tool for the prediction of Gram-negative bacteremia, using cell population data, complete blood count, and differential count. The model seeks to distinguish among nonbacteremia, Gram-negative bacteremia, and Gram-positive bacteremia in patients presenting to the emergency department.

This retrospective study was conducted in the emergency departments of three hospitals in Taiwan. Data from adults with suspected bacterial infections were collected, including complete blood count, white blood cell differential count, and cell population data. A gradient boosting model (Catboost) was developed to classify nonbacteremia, Gram-negative and Gram-positive bacteremia. We evaluated the model through discrimination and calibration.

Here, we show an analysis of 28,503 cases from the China Medical University Hospital developing cohort, including 795 cases of Gram-positive and 2174 cases of Gram-negative bacteremia. Validation cohorts comprise 15,801 cases from China Medical University Hospital, 2632 from Wei-Gong Memorial Hospital, and 3811 from An-Nan Hospital. For Gram-negative bacteremia, the area under the receiver operating characteristic curve ranges from 0.861 to 0.869, with values for the area under the precision-recall curve ranging from 0.325 to 0.415. Predictions for Gram-positive bacteremia are less accurate, with areas under the curve ranging from 0.759 to 0.798 and values between 0.079 and 0.093 for the precision-recall curve.

This study shows that machine learning using hematological parameters provides robust early detection of Gram-negative bacteremia in emergency department settings. Cell population data are valuable predictors by reflecting host immune responses. Data imbalance and marked blood cell changes in Gram-negative bacteria may hinder recognition of Gram-positive bacteremia. Future research should explore the real-world impact of deploying the model in clinical settings.