Research on biomarkers screening for urinary tract infection and the prediction model based on machine learning algorithms.

backgroundUrinary tract infections (UTIs) affect over 1.5 billion people annually, yet systematic screening of cytokine combinations as biomarkers remains limited. This study aims to analyze clinical data with the aid of machine learning algorithms, identify cytokines closely associated with UTIs in our cohort, and explore potential candidate biomarkers that warrant further validation for the early diagnosis, treatment, and prognosis evaluation of UTIs.

methodsClinical data were collected and divided into a training set and a testing set at a ratio of 7:3. During the training phase, the technique of 10-fold cross-validation repeated 10 times combined with grid search was applied to screen out the optimal parameters for five classification algorithms, namely logistic regression, random forest, Adaptive Boosting, Support Vector Machine, and Naive Bayes, and corresponding prediction models were constructed. The final model score was determined by the average accuracy rate over multiple iterations, and a confusion matrix was generated. For each trained and tested model, a Receiver Operating Characteristic (ROC) curve was plotted, and the Area Under the Curve (AUC) was calculated. The model with the best performance was selected through the evaluation of the AUC, and the contribution weights of each feature to this model were analyzed.

resultsAmong the five models, the random forest model demonstrated excellent performance, with its Area Under the Receiver Operating Characteristic Curve (AUROC) reaching 0.836 and the Area Under the Precision-Recall Curve (AUPRC) as high as 0.942. The Support Vector Machine model had the best precision, which was 0.928. The Naive Bayes model stood out in terms of recall and F1-score, with values of 0.840 and 0.841 respectively. Based on the variable importance prediction of the random forest model, it was determined that factors such as age, IFN-γ, IL-6, IL-2, and IL-10 had relatively high importance scores. When evaluating the prediction ability of the models using the testing set, it was found that the AUC and AUPRC values of the random forest and AdaBoost models were significantly higher than those of the other models. The calibration curve analysis further confirmed the high accuracy and reliability of the models.

conclusionsAmong the five models, the random forest model had the best prediction performance. The variable importance analysis of the random forest model identified cytokines such as IFN-γ, IL-6, IL-2, and IL-10 as important features in our cohort, suggesting their potential relevance to the diagnosis of UTIs. These cytokines are expected to become highly potential biomarkers in the precision medicine of UTIs.