Explainable machine learning with routine biomarkers identifies culture-defined bacteremic urosepsis

Area of Science:

• Clinical Medicine
• Infectious Diseases
• Medical Informatics

Background:

• Urosepsis, a severe urinary tract infection (UTI) complication, poses a significant risk of organ dysfunction and mortality.
• Early and accurate identification of bacteremic urosepsis remains a clinical challenge, necessitating improved risk stratification tools.
• Routinely available laboratory data offers a potential avenue for developing predictive models for early risk assessment.

Purpose of the Study:

• To develop and evaluate machine learning models (Random Forest, XGBoost, Logistic Regression) for early risk stratification of bacteremic urosepsis.
• To assess the predictive performance of models using routine biomarkers obtained within 24 hours of patient presentation.
• To identify key routine laboratory predictors for bacteremic urosepsis in hospitalized patients with culture-confirmed UTI.

Main Methods:

• Retrospective analysis of clinical data from 182 hospitalized patients with culture-confirmed UTI.
• Development of Random Forest (RF), Extreme Gradient Boosting (XGBoost), and multivariable Logistic Regression (LR) models using biomarkers from 0-24h.
• Model discrimination assessed via Area Under the Receiver Operating Characteristic Curve (AUC) on a held-out test set.

Main Results:

• XGBoost model achieved an AUC of 0.886, outperforming RF (0.822) and LR (0.822) on the test set, though differences were not statistically significant.
• Key predictors identified included D-dimer, procalcitonin (PCT), C-reactive protein (CRP), white blood cell count (WBC), and albumin.
• Bacteremic urosepsis cases exhibited significantly higher PCT, CRP, WBC, and lower albumin compared to non-bacteremic UTI.

Conclusions:

• Machine learning models, particularly XGBoost, demonstrate good discrimination for identifying bacteremic urosepsis using routine biomarkers within 24 hours.
• D-dimer, procalcitonin, and albumin are significant predictors, highlighting the potential of routine laboratory tests for early risk stratification.
• External validation is recommended to confirm the utility of these models in diverse clinical settings.