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

Area of Science:

• Medical Informatics
• Clinical Microbiology
• Machine Learning in Healthcare

Background:

• Early diagnosis of bloodstream infections (BSI) is critical for appropriate antibiotic stewardship.
• Current diagnostic methods may have limitations in speed and accessibility.
• Biochemical variables offer a potential avenue for rapid BSI risk assessment.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model for early detection of bloodstream infections (BSI) using routine biochemical data.
• To identify key biochemical predictors of BSI.
• To assess the model's performance in a large, real-world clinical dataset.

Main Methods:

• A retrospective analysis of 144,398 patient samples (2010-2020) from Rigshospitalet, Denmark.
• Development of seven ML models, including LightGBM, using demographic and up to 36 biochemical variables.
• Independent testing of the best performing model on 20% of the dataset (10,837 samples).
• Utilized SHapley Additive exPlanations (SHAP) for feature interpretability.

Main Results:

• The LightGBM model achieved an Area Under the Curve (AUC) of 0.69 on the independent test set.
• The model demonstrated high negative predictive value (NPV) of 0.96 and specificity of 0.74, indicating strong performance in identifying patients without BSI.
• Top predictive features included platelets, leukocytes, and the neutrophils-to-lymphocytes ratio.
• Sensitivity for common pathogens like E. coli was 0.71, with an average sensitivity of 0.66 across common BSI pathogens.

Conclusions:

• Biochemical variables hold significant potential as diagnostic factors for early bloodstream infection detection.
• The developed ML model can assist clinicians in identifying patients at low risk for BSI, potentially optimizing antibiotic use.
• Further research and integration into clinical workflows could enhance BSI management strategies.