AI For Clinical Trials All Smoke and Mirrors

By Deborah Borfitz

April 5, 2019 | The problem with applying artificial intelligence (AI) in healthcare is twofold—innovation is antithetical to the corporate culture and data is viewed as the provenance of the IT department, according to Ron Williams, founder and CEO of Business Evolution. For all the talk about AI at industry conferences, most vociferously by vendors in the exhibition hall, the technology has made next to no inroads in the real world, he says.

Williams consults with multiple market segments in and outside of healthcare trying to make sense of data and leverage AI at the operational level. One of his companies, medKyn (Mesopotamian word for “cool medicine”), is devoted to streamlining and automating the logistical processes of clinical research organizations (CROs). In February, he sat on an AI panel at SCOPE, the Summit for Clinical Ops Executives, in Orlando alongside David Vulcano, HCA’s vice president of research compliance & integrity, and Niall O’Connor, chief technology officer for Genospace at HCA’s Sarah Cannon cancer institute.

The onstage banter included jokes about “artificial stupidity” and how “sexy” AI sounds. Williams made the point that 70% of AI projects fail, and that in the clinical trials space it is all “smoke and mirrors.” His research indicates 80% of vendors are making some sort of AI claim that 0% of the time is credible. Often companies are simply putting a form or spreadsheet online with “no way to mine the content to understand what is happening across the enterprise or trial,” he says. Or their tool digs into databases to find more study investigators or participants using machine learning which, technically speaking, is not AI. “More importantly, customers are not getting the value exchange they expect from the application of a so-called AI solution.”

By his definition, there are only two forms of AI—"narrow AI” that does mundane, repetitive tasks very well and can learn to do them better over time, and the fear-inducing “general AI” that could respond dynamically to any situation and won’t exist for at least another 100 years.

Narrow AI or “augmented intelligence” can play a valuable role in healthcare and clinical research, Williams says. But organizations will first need to break some bad habits such as starting with the technology rather than talking to people on the front lines to understand how work gets done or overlooking the time and expense requirements. His recipe for success has a few prerequisites: define the data strategy and what success means, streamline and automate workflow, produce and stack virtual data models, and create a superhighway where data moves from point of origin to point of use.

The current standouts in healthcare, when it comes to AI, are the World Health Organization (WHO) and the National Institutes of Health (NIH), says Williams. They are both well positioned with vast amounts of “secure, aggregated, clean and trusted” data on which they’ve done a lot of analytics. The WHO is working on a repository of use cases of AI in healthcare to identify data formats as well as interoperability mechanisms required to amplify their impact. The NIH is actively exploring ways to incorporate AI into biomedical research and bringing in outside talent to help it better embrace emerging technology.

Across industries, investment in AI is concentrated in the hands of nine for-profit companies—including Google, Facebook and Amazon—and now supersedes that of the federal government. While India and China are pioneering projects for the public good, Williams says, what’s being developed in the U.S. is largely proprietary.

Meanwhile, many of the key players in the clinical research sector continue to rely on spreadsheets and bad data to get products approved with no financial incentive to change, Williams says. AI will one day allow clinical trials with no human beings in them—preventing the approval of dangerous drugs and devices, and the ensuing class-action lawsuits—but it will take a catastrophic event or two for that vision to be realized, he predicts.

The Building Blocks

Getting predictive and prescriptive results from AI requires having the infrastructure allowing an organization to have NIH-quality data, which generally makes the IT department cringe, Williams says. But existing infrastructure need only be surveyed to assess how bad it is, not completely overhauled. IT is in any case not the ideal project lead. Its proper role is to “ensure the security, access, storage and availability of data and to get the right data to the right people at the right time. Everything else must be done by the business owner. That is how you operationalize and monetize data.”

People at the top of the organizational hierarchy must first define what data means to the organization—a raw resource that can be as valuable as diamonds and gold—and reward people for acting accordingly, Williams says. The next step is to identity a problem that needs solving and assign relevant data to it. Breaking the work down into “bite-size chunks,” rather than trying to aggregate all available data at one time, is what makes the process manageable and affordable.

He advises business leaders to build virtual data models around core business processes to understand what data is relevant, where it is and needs to go, and what is supposed to be done with it. This is how national retail brands respond in minutes to common issues at the local store level. Health organizations could similarly be using data analytics to facilitate meaningful, timely action on common problems plaguing the enterprise. The fact that healthcare deals with patients shouldn’t make the processes for dealing with data any different than it for the aviation industry where hundreds of lives are on the line in real time.


Add Comment

Text Only 2000 character limit

Page 1 of 1