By Allison Proffitt

May 3, 2016 | A team of bioinformaticians at Cincinnati Children’s Hospital has applied machine learning to predict which patients will participate in clinical trials. When physicians approached patients, 60% of them agreed to participate in trials. Using an algorithm to predict trial interest, the researchers pushed the acceptance rates to 72%.

The study was published this week in the Journal of the American Medical Informatics Association (DOI: http://dx.doi.org/10.1136/amiajnl-2014-002887).

Clinical trial study coordinators spend quite a lot of time identifying patients for trials, said Yizhao Ni, PhD, lead author and a researcher in the Division of Biomedical Informatics at Cincinnati Children’s, only to be turned down by nearly half of the patients they approach. “So they asked us if there’s any magic that can predict what would be patients response before approaching patients,” Ni told Clinical Informatics News. 

Ni started with literature search to identify the reasons patients decline to participate in trials; patients listed everything from their family and friends’ opinions of clinical trials to characteristics of a specific trial. Age, race, education, socioeconomic level, financial resources and required time commitment are examples of objective factors that influence enrollment. Subjective factors include attitudes about medical research, family influence, seasonality, or whether a person’s health condition has suddenly deteriorated.

“We collected such information and we designed our algorithm,” Ni said. “[Machine learning] can create a mathematic model, which can model the relationship between a target—here the target is whether a patient will enroll in a clinical trial or not—and those factors which [influence their decision].

The goal, Ni said, is to invite patients to participate in trials they will want to be a part of—finding the right trial for a patient. “Patient-directed enrollment is a very new and emerging area,” he said.

It’s a different approach to patient recruitment. This isn’t finding patients for trials, but instead finding the right trial for each patient. “We’ve stopped considering why a patient declines a clinical trial invitation, so we’re trying to study this from a patient’s point of view,” Ni said.

“A patient may be eligible for many clinical trials. In that case we would help researchers pick the clinical trial based on the patient’s preference, based on the patient’s background, and based on the clinical trial’s characteristics.”

To test their algorithm, Ni and his colleagues collected data from 2010 through 2012 involving clinical trial recruitment in the Emergency Department of Cincinnati Children’s. By looking at 18 clinical trials, they found that 6o% of patients approached with traditional recruitment practices ultimately agreed to participate. Researchers predict that their new automated algorithm could push acceptance rates up to about 72%.

“It’s a significant jump, but it’s still a long way from production,” Ni said. His goal is to raise the acceptance rate to 80-90%.

The algorithm won’t help enroll every trial. The study confirmed that patients are less likely to participate in randomized studies, multi-center trials, more complex trials, and trials that required follow-up visits.

“We are still improving our algorithm, but in parallel we are considering moving into a small prospective study that would involve a small group of clinical study coordinators and a small group of clinical trials.”