Algorithm reduces broad-spectrum antibiotic use for UTIs by 67%
A machine learning algorithm reduced the use of broad-spectrum antibiotics for UTIs by as much as 67% in a study conducted at two Boston hospitals, according to results published in Science Translational Medicine.
The algorithm could aid clinicians in mitigating drug resistance by improving prescribing practices for UTIs, researchers said.
“Overprescription of broad-spectrum antibiotics is a major contributor to the problem of antibiotic resistance. Particularly in outpatient settings, providers have few tools to help guide them in choosing the optimal therapy,” David Sontag, PhD, an associate professor of computer science at Massachusetts Institute of Technology, told Healio.
“We’ve addressed this gap by building a machine learning model to predict the probability of antibiotic resistance, and then developed an algorithm that uses these predictions to recommend a treatment,” Sontag said. “It avoids the use of broad-spectrum antibiotics to the greatest extent possible while still choosing an antibiotic that is likely to resolve the patient’s infection.”
Sontag and colleagues created a machine learning model that used data from electronic health records to predict antibiotic resistance to first- and second-line treatments, and an algorithm that recommended “the narrowest possible antibiotic to which a specimen is susceptible.”
The machine learning models were trained using data from 10,053 patients with uncomplicated UTI who were admitted to Massachusetts General Hospital and Brigham and Women’s Hospital from Jan. 1, 2007, to Dec. 31, 2013. The researchers retrospectively tested the algorithm’s performance vs. clinicians using a test set of 3,629 patients with UTI.
The average patient age was 34 years, and 64% of patients were white. Overall, the algorithm cut second-line antibiotic use by 67% relative to clinicians. It also decreased inappropriate antibiotic treatment — defined as a treatment choice that is specimen resistant — by 18%.
When clinicians chose a second-line drug and the algorithm selected a first-line drug, 92% of the time the infections were susceptible to front-line drugs. Additionally, the algorithm chose appropriate first-line drugs 47% of the time when clinicians selected an inappropriate first-line drug.
“Our next step is to conduct a prospective trial to assess the utility of a machine learning treatment recommendation tool embedded into clinical workflows that provides real-time, continually updated, personalized antibiotic treatment recommendations at the point of care as compared to the standard of care,” Sontag said. “We also plan to use the same learning algorithms to provide treatment recommendations for other serious infections such as pneumonia and bloodstream infection, as well as develop ways to estimate the impact of antibiotic treatment on the future risk of resistance.”