Prediction of bloodstream infection using machine learning based primarily on biochemical data

Abstract

Early diagnosis of bloodstream infection (BSI) is crucial for informed antibiotic use. This study developed a machine learning approach for early BSI detection using a comprehensive dataset from Rigshospitalet, Denmark (2010-2020). The dataset included 144,398 samples from adult patients, containing blood culture results, demographics, and up to 36 biochemical variables. Positive blood culture was observed in 6.4% of samples, mostly caused by Staphylococcus aureus, Escherichia coli, and Enterococcus faecium. 80% of the samples (N = 43,351 patients) were used for ML model development and five-fold cross-validation, with 20% for independent testing (N = 10,837). Among seven models, LightGBM performed best, achieving an AUC of 0.69 on the test set. It was more accurate in detecting negatives, with a negative predictive value (NPV) of 0.96 and specificity of 0.74, compared to a positive predictive value (PPV) of 0.13 and sensitivity of 0.54. SHapley Additive exPlanations (SHAP) identified platelets, leukocytes, and neutrophils-to-lymphocytes as the top-3 predictive features. The model showed higher sensitivity (average 0.66) for common pathogens, e.g., 0.71 for E. coli. Results highlight the potential of biochemical variables as diagnostic factors for BSI, indicating clinical use to focus on identifying patients at low risks and can be further enhanced in future investigations.

Keywords: Artificial intelligence; Biomarkers; Classification models; Clinical utility; Diagnosis; Electronic medical records; Infection management; Interpretability; Real-world data.

Fig. 1

Overview of the LightGBM model performance on the test set. (a) Receiver operating characteristics (ROC) and precision-recall curves and confusion matrix for the LightGBM model on the test set where “BSI” denotes confirmed positive blood culture. The numbers in the confusion matrix indicate the number of samples from the test set. (b) Prediction performance of the model, (c) model sensitivity in prediction of top 10 most common pathogens found in the test set.

Fig. 2

SHAP summary plot for the best performing model (LightGBM) from 28,879 predictions on the test set: x-axis is the SHAP value, y-axis is the list of top-20 most contributing features (variables) sorted by their mean absolute SHAP values. Featue values are color coded. Each point indicates an instance (sample). Missing values are indicated by the points in grey color.

Fig. 3

Overview of the analytical pipeline used in this study from data collection and aggregation to machine learning and interpretation.