Predicting Urine Culture Outcomes in Adult Patients Using Machine Learning with the Aim of Reducing Unnecessary Urine Cultures.

BACKGROUND

Urine cultures are frequently ordered tests with a low positivity rate. Development of a machine learning model to predict urine culture outcomes could not only reduce unnecessary urine cultures but also prevent preliminary antibiotic treatment, thereby improving the quality of diagnostic healthcare decision-making.

METHODS

An eXtreme Gradient Boosting (XGBoost) model to predict urine culture outcomes of adult patients was developed. Data was extracted from the electronic health records and laboratory information system of the Amsterdam University Medical Centers (Amsterdam UMC) between 2019 and 2021. Amsterdam UMC is an academic hospital in the Netherlands with 2 separate locations: VU Medical Center (VUmc) and Academic Medical Center (AMC). The VUmc cohort was used for model development and internal validation. External validation was performed in the AMC cohort. All ordering departments were included, i.e., emergency department and inpatient and outpatient clinics. No specific patient groups were excluded.

RESULTS

The VUmc and AMC cohort consisted of 8015 and 10 078 unique urine cultures, respectively. The positive urine culture rate was 19.4% in the VUmc and 12.0% in the AMC. In the VUmc, the model achieved an area under the receiver operating characteristic (AUROC) of 0.834 (95% CI ± 0.010). During external validation in the AMC, the AUROC was 0.800 (95% CI ± 0.015).

CONCLUSIONS

We presented an XGBoost model to predict urine culture outcomes, which retained its performance during external validation. Contrary to most other models, adult patients of all ordering departments were included, which impedes future implementation. An additional external validation and prospective evaluation will be necessary before implementation with the aim of reducing unnecessary urine cultures.