Background: Urinary tract infections (UTIs) impose a substantial burden on hospitals, primarily through antibiotic use. This study aims to assess the extent of empirical antibiotic treatment (EAT) failure, and develop a integrable predictive machine learning model for EAT failure in UTI patients through the use of clinical variables comorbidities, risk factors, and clinical management.
Objectives: To develop models using machine learning algorithms for predicting empirical antibiotic treatment failure in patients with urinary tract infections
Methods: Design & Setting: A retrospective cohort analysis was conducted on 363 hospitalized patients with urinary tract infection in tertiary care hospital in western India Inclusion: Patients were included if their urine culture revealed at least 10^5 colony forming units (CFU)/mL of a uropathogen via positive urine culture. Main Outcome measures: The metrics of accuracy, precision, recall, F1 score, and ROC-AUC curves were used to validate the models. Incidence rate of Empirical Antibiotic treatment failure and severity of treatment failure with patient outcomes, such as sepsis incidence, and mortality. Statistical analysis: A dataset of 48 variables was used to train the models, which included demographic variables, comorbidities, risk factors, resistance data, empirical therapy administered, and outcome variables. The data were split into two sets, a training set, and a test set, at a ratio of 75:25. Seven algorithms were used to train models for treatment failure prediction.
Results: The study revealed a significant empirical antibiotic treatment failure rate of 74.8%, a 34.26% 90-day readmission rate, and a 15.87% mortality rate. Sepsis affected 15.87%, shock was noted in 5.54% patients. Clinical outcomes included a 34.26% rate of readmission within 90 days, 26.45% relapse, and 15.87% re-infection. Sepsis affected 15.87%, shock was noted in 5.54% patients. The gradient boosting classifier had the best test accuracy (78.02%), while the Gaussian naive Bayes classifier had the highest precision (81.13%). The extra tree classifier achieved the best test recall (76.92%) and F1 score (77.78%). The gradient-boosting classifier had the highest AUC at 0.723
Conclusions: The gradient-boosting classifier model exhibited the highest test accuracy and fair discriminatory ability (ROC-AUC score =0.72) among all the models. A comprehensive approach to predicting empirical antibiotic treatment failure using such models can contribute to the growing body of machine learning applications in infectious diseases, healthcare decision-making, and enhancing patient outcomes.
