Ponente
Descripción
Urinary tract infections (UTIs) are a frequent complication in patients with systemic lupus erythematosus (SLE) and may substantially affect quality of life. Previous studies have identified several clinical risk factors associated with infection; however, the interaction among these variables remains incompletely understood. This study proposes a machine learning pipeline to identify relevant risk factors, estimate infection probability, and predict the occurrence of UTI in SLE patients (Cervera, R. et al., Medicine (Baltimore) 78, 167–175, 1999; Hidalgo-Tenorio, C. et al., Ann. Rheum. Dis. 63, 431–437, 2004). The outcome variable was defined as the presence or absence of UTI. Clinical and infectious variables were collected from the study population, and each candidate predictor was evaluated using the Symmetrical Uncertainty (SU) measure (S. Gómez Guerrero G. Sosa Cabrera M. García Torres and C.E. Schaerer. “A Multivariate approach to the Symmetrical Uncertainty Measure: Application to Feature Selection Problem”. In: Information Sciences (2019)). The outcome variable was categorical, and most explanatory variables were also categorical, while a smaller number were numeric and were discretized when necessary. SU was selected because it provides a normalized, nonparametric measure of association between predictors and outcome, with values ranging from 0 to 1. Candidate predictors were ranked according to their SU scores, and the five variables with the highest values were selected as model inputs. Two additional variables were incorporated based on clinical expertise. The dataset included 269 baseline patients and 149 variables, including 25 clinical variables. Following feature selection, predictive models were trained using Logistic Regression, Random Forest, Support Vector Machines (SVM), and Extreme Gradient Boosting (XGBoost). A grouped cross-validation strategy was adopted to prevent information leakage by assigning all observations from the same patient to the same fold. Because positive UTI cases
represented approximately 6% of observations, Synthetic Minority Oversampling Technique (SMOTE) was applied during training to address class imbalance. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and geometric mean (G-mean). Logistic Regression achieved an AUC of 0.977 (95% CI: 0.966–0.985), sensitivity of 0.989, specificity of 0.944, and G-mean of 0.966. Random Forest achieved an AUC of 0.981 (95% CI: 0.972–0.989), sensitivity of 0.735, specificity of 0.977, and G-mean of 0.841. SVM achieved an AUC of 0.968 (95% CI: 0.950–0.982), sensitivity of 0.803, specificity of 0.964, and G-mean of 0.874. XGBoost achieved an AUC of 0.963 (95% CI: 0.924–0.982), sensitivity of 0.702, specificity of 0.969, and G-mean of 0.819. Across all evaluated models, AUC values remained above 0.95, indicating excellent discriminative performance. Interpretability analysis identified previous UTI episodes, age, glucocorticoid therapy, antibiotic therapy, and low fluid intake as the most influential predictors. These variables consistently ranked among the most relevant factors and are supported by previous epidemiological evidence. The results suggest that the proposed framework provides a clinically meaningful approach for identifying risk factors and predicting UTI in SLE patients. Future work will extend the analysis longitudinally and further improve model calibration.