How Machine Learning Can Make Clinical Trials More Successful

Published April 27, 2016
Journal of the American Medical Informatics Association

Scientists are teaching computers how to predict whether people will agree to participate in clinical trials.

Challenges with patient recruitment for clinical trials are a major barrier to timely and efficient translational research,” says Yizhao Ni, PhD, an instructor in the Division of Biomedical Informatics. “The ultimate goal of our research is to increase participation in clinical trials, and to help ensure that studies can be completed with meaningful data.” 

Ni was lead author of a study published April 27, 2016, in the Journal of the American Medical Informatics Association that details the value of machine learning as a potential recruitment aid.

Ni and colleagues gathered demographic information from thousands of emergency department patients and families who were invited to join 18 studies over a three-year period. They also assembled clinical trial data about the complexity and time commitments involved.

The team’s new algorithm had a 72 percent accuracy rate at predicting which patients would agree to join a study. The actual rate, using traditional practices, was 60 percent.

The machine learning algorithm adjusted how to weigh factors including age, race, education, socioeconomic level, attitudes about medical research, and more. It ignored assumptions and biases that might have existed among staff recruiters.

It may be possible to push the algorithm's accuracy beyond 72 percent, Ni says. However, a major challenge will be developing efficient ways to gather the data that automated algorithms need in real time. It remains difficult to process recruitment information in busy medical clinic environments.

Even so, by removing biases, machine learning algorithms eventually could help clinical research coordinators use the data they have to become more effective at matching patients with appropriate clinical studies, Ni says.