When a patient develops a serious bacterial infection, doctors often need to begin treatment before lab results show which antibiotics will work. Researchers at the University of Iowa hope a new machine learning model can make those early decisions more accurate by giving a clearer picture of which drugs are most likely to be effective for each patient.
The study, led by Michihiko Goto, MD, MS, associate professor in Infectious Diseases, details a new way to predict antibiotic resistance using information already available in a patient’s electronic health record. This project was a collaboration with Nick Street, PhD, professor of Business Analytics at the University of Iowa Tippie College of Business, and Anindita Bandyopadhyay, a PhD candidate in business analytics at Tippie. Together, this interdisciplinary team developed a model that estimates how likely a patient’s infection is to resist several different antibiotics.
Today, hospitals use antibiograms—reports that summarize how bacteria have responded to antibiotics across the hospital over the past year—to guide treatment. While those reports are useful, they reflect trends across large groups of patients without factoring in the specifics of an individual patient’s case, which can influence antibiotic resistance. For example, a patient’s previous infections, recent antibiotic use, or other health conditions can affect an antibiotic’s effectiveness.
“Every day, we have to pick an antibiotic before we know what we’re actually treating,” Goto said. “The antibiogram tells us what happened across the whole hospital over the past year, but it can’t tell us anything about the patient in front of us: what infections they’ve had before, what antibiotics they’ve already received, what other conditions they’re managing. The information is already in the medical record. The question is how to put it to work for every prescriber, every time.”
Instead of relying only on hospital-wide data, the researchers trained a machine learning model using millions of pieces of de-identified clinical data, drawn from 127 hospitals and more than 1,400 clinics in 48 states. The model looks for patterns connecting a patient’s medical history to the likelihood that common bacteria will be resistant to different antibiotics. Rather than making one prediction at a time, it evaluates resistance to eight classes of antibiotics simultaneously. This gives physicians a more complete picture when deciding on an initial treatment.
“Two patients can come in on the same day, at the same hospital, with the same organism, and have very different odds of resistance depending on their history,” Goto said.
To test the approach, the team analyzed more than 780,000 samples of Escherichia coli and Klebsiella bacteria collected from Veterans Health Administration hospitals and clinics across the country. After training the model on several years of patient data, the researchers tested it against a full year of more recent data it had never seen before to evaluate how it could perform in real-world use.
The model outperformed traditional antibiograms and standard machine learning methods, particularly when identifying carbapenem resistance—a rare but especially dangerous type of antibiotic resistance that can leave patients with few treatment options.
The project brought together expertise from two different fields. Goto’s background in infectious disease and antimicrobial stewardship helped define the clinical questions, while Street contributed his experience in predictive modeling and machine learning.
Although the model was developed using two of the bacteria most commonly responsible for serious infections, the researchers see this as an important first step. Future studies will expand to other organisms and explore how personalized antibiograms could fit into everyday clinical practice.
“We’ve shown this can work for the two commonly seen organisms, and that’s a starting point rather than an endpoint,” Goto said. “The longer-term goal is a personalized antibiogram built directly into the electronic health record, updating as a patient’s history changes, so the information arrives at the moment the decision is being made rather than in a report published a year later. This isn’t meant to replace clinical judgment. It’s meant to give clinicians better information at the moment they have to act.”