Can machine learning algorithms predict negative urine cultures using flow cytometry routine data?

Introduction

Urine samples are among the most frequent samples analysed in microbiology laboratories, while a large proportion are culture-negative. The aim of this study was to test whether culture-negative samples can be predicted from routine flow cytometric data and how reliable prediction algorithms are over time.

Materials & Methods

In 2023, 1325 urine samples from the University Hospital Münster were used for a training dataset (n=1032) and three independent test datasets (n=93-100 samples) that were collected three months apart. Predictors from flow cytometry were total counts per µl of bacteria, erythrocytes, yeast-like-cells, hyaline casts, crystals, leukocytes, squamous epithelial cells, non-hyaline casts and non-squamous epithelial cells in addition to type of urine sample as well as age and sex of the patient. Two different labels were applied (positive culture defined as any bacterial growth [yes/no] and detection of a clinically relevant uropathogenic species [yes, no]). Three classifiers (decision tree, random forest classifier, CatBoost) were cross-validated on the training dataset with 5 folds in order to select an optimized model with at least 95% sensitivity and maximum negative predictive value (NPV). Finally, the optimized model was trained on the complete training dataset and then evaluated on the three independent test sets.

Results

In total, 72.5% (960/1325) samples were culture positive with a predominance of Escherichia coli (n=295). In all three test sets the classifier predicted a negative culture with a moderate "balanced accuracy" (63-69%), "sensitivity" (94-97%) and "specificity" (31-44%). The test performance of the prediction algorithms was stable over a period of six months. Depending on the tested samples, the negative predictive value was 73-85%.

Summary

The machine learning algorithms in our study population had a moderate NPV to rule out culture negative isolates. Additional clinical parameters (e.g. pH) might improve the test performance.