Q&A: Novel UTI detection method may help curb antibiotic overprescribing in primary care

A new method of UTI detection designed specifically for use in primary care predicted with high accuracy when antibiotics could be safely withheld from low-risk patients, according to researchers.

Daniel Parente, MD, PhD, FAAFP, an assistant professor of family medicine and community health and research director of the family medicine residency program at the University of Kansas Medical Center, and colleagues wrote that UTI symptoms are common in primary care, but results from the reference standard test for infection — a urine culture — take more than a day. These delays can lead to antibiotic overprescribing, fueling ongoing concerns regarding antibiotic resistance.

To improve prescribing practices, the researchers redesigned a urine culture prediction model — the NeedMicro classifier — that previously demonstrated success in an ED setting so that it could be feasibly implemented in primary care. To do this, they removed urine microscopy features that are not often available in primary care and found that removing these features “did not severely compromise performance under internal validation.”

Parente spoke with Healio about the redesigned model, known as the NoMicro classifier, how it compares to traditional UTI detection methods, the impact it could have on practices and more.

Healio: Can you briefly describe the NoMicro classifier?

Parente: The NoMicro classifier uses machine learning to predict whether patients attending a primary care appointment with symptoms of a urine infection will have a culture that grows pathogenic bacteria. Without the NoMicro classifier, clinicians would have to wait for up to 24 hours to get culture results, but decisions about antibiotics are often made right away, before the cultures are available. We used data that had been previously published by a different research group that built a urine culture predictor for use in emergency departments. However, this classifier needed to have information about urine microscopy — cells and bacteria seen under the microscope — to make predictions, which many primary care offices cannot check at the point of care. NoMicro is different in that it avoids needing to use information about urine microscopy and relies on only information that primary care providers usually would have available.

Healio: How does it work? How effective is it?

Parente: The NoMicro classifier is a machine learning software program. It takes as input information about several clinical features, including demographics (age, gender), urine dipstick results (nitrites, leukocytes, blood, clarity, glucose, protein), symptoms (abdominal pain, urinary burning), and whether a patient had urine infections in the past. It uses this to calculate a numerical “score” that a urine culture will end up growing pathogenic bacteria. The score gets compared to a “cutoff” number, which lets it convert the score into a yes-or-no prediction of “likely pathogenic” or “likely not pathogenic.” By adjusting the cutoff, you can tune machine learning predictors to help with specific clinical situations. In our case, we were interested in trying to reduce antibiotic overuse. We wanted to be sure that predictions of “likely not pathogenic” were very reliable. After tuning our model for this, it achieved a negative predictive value of 92.3%.

Healio: How does it compare to traditional methods of UTI detection?

Parente: Machine learning-based methods are major improvements over traditional methods of UTI detection, like relying on just urine dipstick information. Machine learning methods still need that information (urine dipstick) but can also notice patterns in symptoms and urine testing that simpler methods just cannot account for.

Healio: What kind of impact could this have on the issue of antibiotic overprescribing and overuse?

Parente: We hope that the ability to predict when cultures will be negative (normal) might allow providers to reduce antibiotic overuse. We used historical data to evaluate this potential. Physicians are already pretty good at deciding when patients need antibiotics — even without the culture results — but our data suggests they can do even better using our method. However, we will need to do more work before clinicians can safely use our classifier on real patients. All of our work used historical data. We will need to perform prospective clinical trials to make sure that the benefits of reducing antibiotic overuse are not outweighed by any risks.

Healio: What are the clinical implications of your findings?

Parente: We still need to do some further clinical testing on our algorithms before primary care doctors start routinely using our predictor in real-world practice. The most important thing we need to do is to make sure that it is safe to not prescribe antibiotics when the model thinks a patient is likely to have a normal culture. We think it probably is safe, based on the historical data we used in our study, but we can only be sure once we do a prospective clinical trial. We hope that, if those clinical trials go well, that physicians will be able to confidently use our method to reduce antibiotic overuse.

Healio: What is the take-home message for primary care providers?

Parente: Artificial intelligence and machine learning are quickly expanding our ability to recognize patterns in complicated situations. Primary care clinicians are likely to benefit from this technological revolution. Our application of machine learning to urine culture prediction is just one example of how this kind of technology can be used in primary care.

Healio: Is there anything else you’d like to add?

Parente: We’d love to connect with practicing physicians or researchers who are interested in how artificial intelligence and machine learning will change the practice of primary care.

References:

• Dhanda G, et al. Ann Fam Med. 2023;doi:10.1370/afm.2902.
• Taylor RA, et al. PLOS One. 2018;doi:10.1371/journal.pone.0194085.

For more information:

Daniel Parente, MD, PhD, FAAFP, is an assistant professor of family medicine and community health and research director of the family medicine residency program at the University of Kansas Medical Center. He can be reached at dparente@kumc.edu.