Advertisement
Up to 3 days faster than waiting for urine culture results
Adaptable and dynamic computerized algorithms can accurately select antibiotics for urinary tract infection (UTI) up to three days before urine culture results are available.
Advertisement
Cleveland Clinic is a non-profit academic medical center. Advertising on our site helps support our mission. We do not endorse non-Cleveland Clinic products or services. Policy
This was the chief finding from a study presented by Cleveland Clinic investigators at the American Urological Association (AUA) 2024 meeting. The study tested trained algorithms on about 6 million patient cases of resistance to 11 of the most clinically relevant first-line antibiotics used to treat UTI.
“This research is novel and very exciting because no one has previously developed algorithms for antibiotic-resistance prediction using this large of a training dataset or with the level of accuracy that we’re seeing,” says physician-scientist Glenn Werneburg, MD, PhD, lead author of the study and a resident in Cleveland Clinic’s Department of Urology. “We believe these tools can improve patient outcomes, reduce time to resolution of symptoms and improve antibiotic stewardship at the population level.”
Urine culture is the standard for diagnosing UTIs, but it can take up to 72 hours to receive the results. During this time, clinicians often treat patients presumptively, based on symptoms and risk factors. In many cases, those initial drug choices are proven wrong by sensitivities identified on cultures, leading to a switch in therapy.
The goal of the new research was to create machine learning algorithms that accurately predict UTI drug sensitivity before culture results are available, decreasing time to symptom resolution, minimizing the need for prolonged antibiotic therapy, and reducing patient burden and clinician workload.
Dr. Werneburg and coauthors developed a machine learning algorithm for each of the 11 antibiotics, including vancomycin and nitrofurantoin. Algorithms incorporated a series of known and putative risk factors for antibiotic resistance. The culture data on which the models were trained were from Cleveland Clinic’s electronic medical record (EMR), collected across the health system over seven years. Through the EMR, urine cultures with known drug sensitivities were identified.
Advertisement
Overall, the training and validation process involved more than 560,000 cultures and testing of the algorithms on about 6 million cases of antibiotic sensitivity/resistance. An area under the receiver operating characteristic curve (AUC) for the automated predictions by each of the algorithms was calculated.
The models that performed best were ensemble algorithms — machine learning that combines multiple individual models for more robust prediction. For most of the antibiotics tested, those types of algorithms had an AUC of 0.75 for forecasting drug sensitivity before culture data were available. The AUC for prediction of resistance to vancomycin, for example, was 0.81 (accuracy 0.73, precision 0.70, recall 0.82).
The authors used a drop-out method to identify the most relevant factors in each model. The algorithms are designed to iteratively update and thus become more accurate once an organism’s identity (but not yet antibiotic sensitivity) becomes available from a culture. For nitrofurantoin, for example, the AUC for resistance prediction rose to 0.93 once an organism had been identified in a culture.
“One of the most encouraging aspects of the study is that the algorithms were accurate when we validated them in our hospital and also in an external dataset at a geographically distant hospital,” says Dr. Werneburg. “That bodes well for implementation at other hospitals nationally and internationally.”
The next step for the researchers is to improve the mathematics underlying the algorithms and incorporate novel aspects of machine learning to further refine the calculations. They also plan to compare the performance of the tools with the performance of clinicians. The ultimate goal is to improve care of patients with UTIs.
Advertisement
“Our results have important implications for the reduction of antibiotic resistance, through more accurately targeted therapy for UTI,” says Cleveland Clinic urologist Sandip Vasavada, MD, senior author of the study. “We believe that clinical use of the algorithms will ensure that patients with these very common infections get optimal treatment more quickly and at a lower cost to the healthcare system.”
Advertisement
Advertisement
Cleveland Clinic uses data to drive its AI implementation strategy
Pairing machine learning with multi-omics revealed potential therapeutic targets
Cleveland Clinic and IBM leaders share insights, concerns, optimism about impacts
Scientific program chair reflects on what may resonate longest from this year’s neurosurgery conference
Customized bots improve speed, efficiency by streamlining daily clinical, clerical tasks
Cleveland Clinic neuromuscular specialist shares insights on AI in his field and beyond
Neurology is especially well positioned for opportunities to enhance clinical care and medical training
How technology is changing how and whom we hire