Much of the increases in healthcare costs can be attributed to the rise in price of pharmaceuticals, and one reason pharmaceutical expenses are so high is because clinical trials are long, costly, and difficult to run. SMU and UTSW researcher Vishal Ahuja and his colleague from the University of Chicago John R. Birge hope to use statistics to change the way trials are performed, saving time, money, and improving outcomes for patients.
Pharmaceutical companies have to run clinical trials to test new medicines, but they can take years with hundreds of patients, who are difficult to recruit and retain. According to a report by the Manhattan Institute, more than 90 percent of a drug’s development costs are from Phase III clinical trials, where people are recruited to test the drug. Ahuja says that trials can cost between $300-$600 million.
The length and cost of the trial are a result of the way clinical trials are traditionally performed, where half of the subjects are given the medicine and the other half are given a placebo or the existing treatment. That half-and-half ratio is kept all the way through the trial, even if the new drug is showing signs of being more or less effective. In the 1970s, Dr. Don Berry at M.D. Anderson in Houston began pioneering adaptive models for clinical trials, where results from the trial are used to determine how many people receive the next round of the tested medicine. If the drug appears to be effective, more people will be randomly assigned to take the medicine, giving researchers more data that will lead to a definitive opinion about the drug.
Adaptive trials have the capability to reduce the length of trials, but require sequential work with just one person at a time, because changes are being made to the ratios as results of the medicine come in. The adaptive trial must consider all the possible scenarios of how a person might react to a drug over time, resulting in billions of combinations that can be a computational nightmare when making decisions about the new treatment.
This is where Ahuja’s work comes in.
Ahuja is using a Bayesian model to reduce the computational load of the adaptive trials. Rather than considering billions of combinations of patient outcomes, his theory groups patients who react similarly into blocks, bringing down the number of scenarios to consider. This Bayesian model works like an Amazon algorithm that predicts what you may want to purchase based on past purchases. Each shopper is unique, but Amazon makes assumptions based on past results about what would be an effective product to pop up on the screen. Ahuja’s work is a similar calculation about patients in a clinical trial, creating a shortcut for those performing the trial. Just like the “you may also like” section saves you shopping time, the Bayesian model can get more patients on an effective medicine or off an ineffective one sooner than traditional trials, while reducing the time it takes to get an answer about the drug.
When Ahuja ran his model against an existing nausea medication trial, it would have reduced the failed patients who didn’t receive an effective drug by 17 percent, and would have shortened the trial by over 20 percent, or about 18 weeks.
Adaptive models are becoming more and more common, but the computations required make them difficult to perform. This new Bayesian model can help reduce cost, time, and increase effectiveness. While many clinical trial operators have shown resistance, academic centers like UT Southwestern are more receptive to these models. “The doctors are also researchers, and they have the understanding and nuance to implement those trials,” he says. He is hopeful that adaptive models will become more common in the future.
The research is described in a paper, An Approximation Approach for Response Adaptive Clinical Trial Design is by Vishal Ahuja, Southern Methodist University’s Cox School of Business, and John Birge of University of Chicago Booth School of Business is currenlty under review at INFORMS Journal on Computing. The entire paper can be found here, and a summary can be found here.