Multifactor machine learning models for predicting urinary tract infections: a pilot study

Purpose: Vitamin D, a fat-soluble prohormone essential for calcium–phosphate homeostasis and bone health, also regulates innate and adaptive immunity through receptors expressed on B cells, T cells, and antigen-presenting cells capable of synthesizing its active form. Deficiency in vitamin D is linked to dysregulated immune responses and an increased risk of autoimmune diseases and infections, particularly urinary tract infections (UTIs) in both children and adults. Here, we explore 12 machine learning models that utilize urinary 25-hydroxyvitamin D (25(OH)D) levels, urine pH, gender, and age to predict UTIs. Methods: A cohort of 358 subjects was analyzed. Demographic, biochemical, and microbiological data were collected for each participant. The dataset was randomly divided into a training set (70%) and an independent test set (30%). Four predictors, age, gender, urine pH, and vitamin D, were included in the analysis. Twelve machine learning models were assessed based on accuracy, specificity, sensitivity, positive predictive value (PPV), negative predictive value (NPV), area under the ROC curve (AUC-ROC), and F1 score. Results: A significant difference in urinary 25(OH)D levels was found between individuals with positive (1.33 ± 4.13 ng/mL) and negative (2.48 ± 4.52 ng/mL) urine cultures (p < 0.001). Using urinary 25(OH)D, urine pH, age, and gender as predictors, 12 machine learning models showed accuracies of 64–87%, sensitivities of 59–79%, specificities of 51–95%, PPVs of 61–94%, NPVs of 63–82%, AUC-ROC values of 0.63–0.93, and F1 scores of 0.63–0.86. A stacking machine learning model achieved 88% accuracy, 83% sensitivity, 94% specificity, 93% PPV, 84% NPV, AUC-ROC of 0.93, and an F1 score of 0.88. Conclusion: Significant differences in urinary 25(OH)D levels between positive and negative urine cultures confirm the association between low vitamin D levels and UTI occurrence. The developed machine learning models demonstrated high accuracy and represent a promising adjunct for clinicians in UTI diagnosis. With additional validation and assay development, such models may eventually complement conventional culture methods in clinical screening programs. Further external validation using independent datasets, along with prospective studies assessing their impact on antibiotic prescribing practices, is warranted. While these models estimate UTI risk, they do not identify the causative pathogen or determine antibiotic susceptibility.