Machine Learning Bests Experts in Predicting Patient Response to Overactive Bladder Treatment

Algorithm aligns with subjective patient reports as well as objective outcomes

Novel machine learning algorithms can be used to assist clinicians in predicting patient-reported outcomes following third-line treatments for overactive bladder (OAB).

Advertising Policy

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

Characterized by symptoms of urinary urgency, frequency and incontinence, OAB is a common and costly urologic diagnosis, affecting about 16% of the U.S. population.

Many patients respond to behavioral approaches or oral medications. But for those who don’t, third-line options include bladder injection of OnabotulinumtoxinA (OBTX‐A) and sacral neuromodulation (SNM). Since those treatments are invasive, it’s important to know in advance which individual patients are more likely to respond to either, and if so to which one.  

Previously, machine learning algorithms developed by Cleveland Clinic researchers using neural networks accurately predicted the objective outcome of reduction in urge urinary incontinence episodes with OBTX‐A and SNM in patients with OAB who had been refractory to conventional therapies. The algorithms’ predictions were at least as accurate as those of two expert urologists, and in some cases better. Those findings were published in March 2022 in Neurourology and Urodynamics.

In a new analysis, the algorithms also predicted patient-reported outcomes in bladder leakage and function, which don’t necessarily correspond to objective measures. The findings were presented in May 2022 at the American Urological Association meeting by the study’s lead author Glenn T. Werneburg, MD, PhD, a urology resident at Cleveland Clinic. The study was later published in the International Urogynecology Journal.

“We were interested in using machine learning-based approaches to predict the patients who would perceive versus not perceive improvement in their symptoms following treatment. If we can predict who’s not going to respond we can more carefully target a specific treatment to a specific patient. This may lead to improved outcomes for our patients, and reduced costs both to the patient and the healthcare system,” Dr. Werneburg says.

The new algorithms were created to predict improvement in bladder leakage and function measured by the validated Patient Global Impression of Improvement (PGI-I) questionnaire. As with the prior study, the algorithms were applied to the data from ROSETTA, a trial sponsored by the National Institute of Child Health and Development in which 381 women with refractory OAB were randomized to OBTX‐A or SNM.

Advertising Policy

Sandip Vasavada, MD,  a female pelvic medicine and reconstructive surgeon at Cleveland Clinic, was an investigator in the ROSETTA trial and also senior author of these machine learning and OAB studies. He says the ROSETTA trial enabled access to secondary data and endpoints that have been essential in these investigations. 

He says, “We were able to leverage these datasets and have access to randomized, controlled, blinded data and then use it for identification of these types of parameters.”

In the design of this study, blinded expert urologists were given the same training data, and this time tasked to predict patient-reported outcomes.

The top algorithms showed excellent predictive accuracy for patient-reported bladder function improvement for both OBTX-A (Area under the curve 0.86) and SNM (AUC 0.96) and were noninferior to expert urologists. The algorithms were also highly accurate in predicting patient-reported bladder leakage improvement for both modalities (AUC 0.75 for OBTX-A and 0.80 for SNM) and were noninferior to experts.

Machine learning complements, but doesn’t replace clinical judgement

Despite their accuracy, the algorithms won’t replace clinician expertise, says Dr. Werneburg, “Some aspects of the physician‐patient interaction are subtle and uncomputable, and thus machine learning may complement, but not replace, a physician’s judgment,” he says.

The next step, Dr. Werneburg says, is to conduct a prospective analysis that would give the clinician, not the algorithm, the benefit of the doubt. “We want to validate these results in a prospective manner wherein we give the clinicians the chance to ask the patients what they think is important and to direct their questions based on the clinical picture and the patient’s history. And then alongside that, task the algorithms with making the same prediction.”

Advertising Policy

They’ll also be looking to see how well the algorithms may be generalizable to men with OAB, he notes.

Ultimately, the idea would be to generate a computer interface where both the clinician and the patient can enter information to inform the prediction of response to one or the other treatment. Such a tool could have widespread use, he noted.

“We suspect that our results are not only applicable to overactive bladder or even urology for that matter but really any context where a decision based on data is necessary. It might help direct patient counseling. Of course, some patients have preferences for one treatment over another, and, in some cases, one treatment may be indicated over another, so we don’t anticipate—even after all the validation—that these will take the place of the doctor visit or the clinician-patient interaction but, really, will just complement it,” says Dr. Werneburg.

Dr. Vasavada concurs, adding, “There may be a patient subgroup, either defined by age or bladder testing parameter— any number of variables—that might enable us to advise patients on potential risk. At the end of the day, the patient decides; but this information might help guide patient decision-making and help them make a better, more informed decision.”