Abstract

Objectives

Urinary tract infections are a common diagnostic challenge. Although urine culture remains the gold standard, it is time-consuming and often ordered reflexively. This study aimed to develop and validate an interpretable machine-learning–based Laboratory Decision-Support System (LDSS) to guide reflective urine culture prioritization using only structured laboratory data.

Methods

We analyzed a retrospective cohort of 51,923 adult patients. Seven machine learning algorithms were trained, with the random forest (RF) model demonstrating the highest accuracy. SHapley Additive exPlanations (SHAP) analysis was employed to ensure model interpretability. A reduced RF model, using the top 10 predictive features, was used to construct three scoring systems: one emphasizing model fidelity, one optimizing diagnostic balance, and one maximizing sensitivity.

Results

The RF model demonstrated excellent performance (external ROC-AUC: 0.956). The simplified 10-variable model maintained high accuracy (ROC-AUC: 0.947). Key predictors included bacterial count, leukocyte count, nitrite presence, and patient age. The scoring systems offered flexible options tailored to different diagnostic priorities, with the SAFE-Score achieving 95.3% sensitivity.

Conclusion

The LDSS is intended to support reflex culture prioritization, not reduce overall culture testing. By streamlining pre-analytical triage and highlighting clinically significant samples, it promotes appropriate culture utilization and strengthens antimicrobial stewardship, while preserving the central role of urine culture in infection management.