Overheard: Randomized Trials and Big Data Get Hitched

Spring 2016

Derek Angus, an MD/MPH, has long argued that multicenter megatrials result in the most informed clinical care—especially for complex conditions such as the brain cancer glioblastoma multiforme, pneumonia, or sepsis. Based in part on a blockbuster study he led, a new definition of, and treatment guidelines for, sepsis appeared in the February 23 issue of JAMA. (Read our Fall 2014 cover story to learn more about sepsis and other megatrials.)

But randomized clinical trials are long and expensive, and they can be discomforting for patients and clinicians. In a 2015 JAMA opinion piece, Angus made the case for a new kind of randomized trial that “fuses” with big data. “The big problem with big data is that there’s no randomization in it, and the singular beauty of randomization is that you can gain causal inference,” he says. Marrying them, says Pitt’s Angus (who is a Distinguished Professor, the Mitchell P. Fink Professor, and chair of critical care medicine), “is an idea whose time has come.”

What’s the fusion you’re proposing?

With the rise of big data and electronic health records (EHR), a number of groups have suggested that you could essentially leverage the EHR to create live estimates of the likelihood of getting benefit from a treatment by running, essentially, a large observational cohort study inside the EHR. But we propose going further: using clinical data in the EHR to influence the ongoing trial.

How is this playing out in the clinic?

We’ve received funding from the European Union and the National Health and Medical Research Council of Australia to launch this program in severe pneumonia patients coming to the ICU. We will be testing multiple antibiotic strategies, whether to immunomodulate the patient with corticosteroids, and . . . different ways of providing mechanical ventilatory support—all at the same time. All generate separate weights of evidence and separate probabilities for different subgroups of patients with pneumonia, depending on how bad their oxygenation is and whether they have shock or not. The trial is simultaneously generating 48 separate measures of treatment effects, as opposed to a single normal trial that generates one. And if any particular combination of therapies in any particular patient subgroup is doing better than the others, the next patient who presents will have the odds weighted in his or her favor toward the best performing therapy. So the trial is constantly learning. We are incredibly excited about envisioning a future where clinical care becomes a constant learning tool.