Personalizing the empiric treatment of gonorrhea using machine learning models.

Abstract

Despite the emergence of antimicrobial-resistant (AMR) strains of Neisseria gonorrhoeae, the treatment of gonorrhea remains empiric and according to standardized guidelines, which are informed by the national prevalence of resistant strains. Yet, the prevalence of AMR varies substantially across geographic and demographic groups. We investigated whether data from the national surveillance system of AMR gonorrhea in the US could be used to personalize the empiric treatment of gonorrhea. We used data from the Gonococcal Isolate Surveillance Project collected between 2000-2010 to train and validate machine learning models to identify resistance to ciprofloxacin (CIP), one of the recommended first-line antibiotics until 2007. We used these models to personalize empiric treatments based on sexual behavior and geographic location and compared their performance with standardized guidelines, which recommended treatment with CIP, ceftriaxone (CRO), or cefixime (CFX) between 2005-2006, and either CRO or CFX between 2007-2010. Compared with standardized guidelines, the personalized treatments could have replaced 33% of CRO and CFX use with CIP while ensuring that 98% of patients were prescribed effective treatment during 2005-2010. The models maintained their performance over time and across geographic regions. Predictive models trained on data from national surveillance systems of AMR gonorrhea could be used to personalize the empiric treatment of gonorrhea based on patients' basic characteristics at the point of care. This approach could reduce the unnecessary use of newer antibiotics while maintaining the effectiveness of first-line therapy.