Using machine learning, researchers have developed a treatment algorithm that could help improve antibiotic prescribing for uncomplicated urinary tract infections (UTIs), according to a study today in Science Translational Medicine.
UTIs are one of most common conditions for which antibiotics are prescribed in the United States, resulting in 4.7 million prescriptions annually. But in more than 40% of cases of uncomplicated UTI, clinicians prescribe fluoroquinolones, which are the second-line therapy according to guidelines from the Infectious Diseases Society of America (IDSA). The first-line antibiotics are the narrow-spectrum nitrofurantoin and trimethoprim-sulfamethoxazole.
The decision to use broader-spectrum fluoroquinolones is likely related to concerns about rising antibiotic resistance to first-line treatment for UTIs. And in emergency rooms and other outpatient settings where UTIs are diagnosed and patients are sent home with an antibiotic, clinicians may favor empiric treatment with agents like ciprofloxacin or levofloxacin to minimize the risk of first-line therapy failure.
But this is problematic, because fluoroquinolones are associated with adverse events like tendon rupture and peripheral neuropathy, and increased use of fluoroquinolones can increase the risk of Clostridioides difficile infections in patients and promote the emergence of multidrug-resistant organisms.
Predicting the probability of resistance
To address this problem and provide outpatient clinicians with a tool to help identify the appropriate antibiotic for UTI patients, a team led by researchers from Harvard Medical School and the Massachusetts Institute of Technology developed a clinical decision support tool that takes data from electronic health records to predict the probability of antibiotic resistance to first- and second-line antibiotics for UTIs. They then developed an algorithm that translates those probabilities into recommendations designed to select the narrowest-spectrum antibiotic that would still be effective.
"The key tension faced by all healthcare providers faced with having to treat an infection is how to choose the narrowest-possible antibiotic, without overdoing it by selecting one to which the microorganism may be resistant, thus risking treatment failure," lead author Sanjat Kanjilal, MD, MPH, an infectious diseases physician at Brigham and Women's Hospital in Boston and a lecturer at Harvard Medical School, said in a video.
Kanjilal and his colleagues built the machine learning model using data from 10,053 women treated for uncomplicated UTIs at Massachusetts General Hospital and Brigham and Women's Hospital from 2007 through 2013 and trained it to predict the probability of antibiotic resistance to nitrofurantoin, trimethoprim-sulfamethoxazole, ciprofloxacin, and levofloxacin.
Next, they tested the algorithm on data from a cohort of 3,629 women treated for uncomplicated UTIs at the two hospitals from 2014 through 2016 and compared the performance with empiric treatment decisions made by clinicians.
The results showed that the algorithm recommended ciprofloxacin or levofloxacin in 11% of specimens from the women, a 67% reduction in the use of those antibiotics compared with recommendations from the clinicians. The algorithm recommended inappropriate antibiotic therapy—defined as an antibiotic to which a specimen is resistant—in 9.8% of cases, an 18% reduction compared with clinicians.
Compared with best-case implementation of the IDSA treatment guidelines, the algorithm had a similar rate of inappropriate antibiotic recommendations (10.7%).
Kanjilal and his colleagues say they hope that the algorithm could ultimately be embedded in the electronic health record and used at the point of care by clinicians. They also want to see if they can apply the algorithm to other bacterial infections to optimize antibiotic treatment.
"This study is one of the first to show that machine learning models can be used for antimicrobial stewardship, and our hope is to expand on this work by trying this approach with other infectious syndromes, such as bloodstream infection and pneumonia," he said.
