Development and validation of a diagnostic model of antimicrobial resistance by extended-spectrum beta-lactamases in community-acquired urinary tract infections. How to adjust the prediction in variable outcome prevalences? The usefulness of contracting predictors with LASSO regression

Abstract

Background: In the case of community-acquired urinary tract infection, the identification of Enterobacteriaceae with extended spectrum beta-lactamases (ESBL) can optimize treatment, control and follow-up strategies, however the effect of variable prevalences of this resistance pattern has affected the external validity of this type of models. Aim: To develop a diagnostic predictive model that adjusts the prediction error in variable prevalences using the LASSO regression. Methods: A diagnostic predictive model of community-acquired urinary tract infection by extended spectrum beta-lactamase-producing Enterobacteriaceae (ESBL) was designed. A cross-sectional study was used for both construction and validation. To assess the effect of the variable prevalence of the outcome, the validation was performed with a population in which the proportion of isolates with this resistance mechanism was lower, the participants were adult patients who consulted the emergency services of two medium-level hospital institutions. complexity of the city of Medellin. To adjust for the effect of an environment with a lower proportion of antimicrobial resistance, we used the contraction of predictors by LASSO regression. Results: 303 patients were included for the construction of the model, six predictors were evaluated and validation was carried out in 220 patients. Conclusion: The adjusted model with LASSO regression favored the external validity of the model in populations with a proportion of ESBL isolates in urine culture of outpatients between 11 and 16%. This study provides criteria for early isolation when predictors are present in populations with proportions of resistance in ambulatory urine cultures close to 15% and proposes a methodology for the adjustment of errors in the design of prediction models for antimicrobial resistance.