Skip Navigation


Looking For TroubleMichael Morgenstern

Looking for Trouble (continued)

From Claims to Crystal Balls

Twenty-five years ago, Weiner and the late Barbara Starfield, MD, MPH ’63, began dreaming up a tool that would act as the next best thing to a crystal ball. It would squeeze health data from computerized insurance claims to categorize patients by risk. This would help highlight patients likeliest to require expensive care in the future. The result of their efforts is the Johns Hopkins Adjusted Clinical Groups (ACG) Case-Mix System.

The flow of information into the ACG system is simple, at least in concept. Each time a patient receives a diagnosis—and there are over 75,000 of them—it’s coded by five criteria: duration, severity, diagnostic certainty, treatment type, and specialty care needed. Each patient’s diagnoses are viewed in combination to slot them into an overarching risk group, or an “adjusted clinical group.” Predictive algorithms then kick in to gauge the likelihood of future expenses or events such as unanticipated hospitalizations.

At first, Weiner, Starfield, and their collaborators simply gave the ACG software away. Yet it’s proven so useful that the School has partnered with six IT corporations around the globe to distribute the system.

ACG now touches more than 80 million people in 17 countries. In terms of revenues paid to Hopkins, this represents the largest tech transfer in the history of the university.

Now, EMRs, along with other “e-health” data sources such as websites and smartphone apps, are poised to take such predictive modeling to an entirely new level, says Weiner. But the strength of these new digital data—that they contain a staggering wealth and diversity of health information—is also what makes this a challenge.

One major problem, says Weiner, is prose.

Even in the best EMR systems, key information is usually embedded in free-text notes, which can take staff precious hours to read. Like it or not, Weiner says, “many doctors just use EMR systems as fancy typewriters” to jot down their thoughts.

Rather than fight human nature by requiring doctors to use drop-downs and radio buttons, Weiner recruited colleagues in “natural language processing,” a field of computer science that involves culling useful information from prose.

One of Weiner’s new collaborators is Mark Dredze, who spent time at Google and is now an assistant research professor of Computer Science at Johns Hopkins. You know how Google finds exactly what you want, no matter how complex the search, asks Dredze. That’s what ought to be possible when programs peruse health care records. Weiner, Dredze, and other interested faculty have taken on finding JHHC’s “missing patients” as a pilot project.

The team began with a field trip to JHHC’s East Baltimore campus, where a nurse walked them through sample health records, pointing out warning signs and, equally important, linguistic permutations. “We might see in the text that this patient is a tobacco user,” explains Dredze. “We also might see ‘former tobacco user’ or ‘this patient is not a tobacco user’ or ‘this patient lives with a tobacco user.’” Distinguishing between such phrases requires a program that can “do what you can think of as diagramming sentences,” says Dredze, not to mention “learning” that smoking is an important trait in the first place.

The project is still in its early stages. (Some of its initial triumphs involved clearing bureaucratic hurdles, such as obtaining permission to access and link three types of electronic records—OB and primary care charts, and insurance files—a first for Weiner.) Looking ahead, both he and Dredze are confident; it’s a small project with a concrete objective, but one that has implications for every major health system in America, including Johns Hopkins, which is in the process of investing hundreds of millions of dollars in a new state-of-the-art EMR system.

". . . many doctors just use EMR systems as fancy typewriters."—Jonathan Weiner

Tinkering Together

Hoping to tackle more real-world challenges in the future, Weiner launched CPHIT last year. He envisions CPHIT as a place where academics from across the University, private and public health care organizations and e-health companies can work together to do R&D and tinker on innovative projects for the common good—a public health version of MIT’s famed Media Lab.

As potential collaborations emerge, Weiner builds a team and gets rolling. CPHIT has begun working with Maryland’s health information exchange to flag patients who are at high risk for re-hospitalization. (Not a bad idea, now that many payers have begun penalizing hospitals for high rates of readmission.) CPHIT is also working with several HMOs, among other organizations, to find better ways to treat chronic diseases. For example, using height and weight, patterns of care seeking, and behavioral risk factors to prioritize outreach or interventions for diabetes patients. And CPHIT is in the early stages of discussions with Sharfstein about developing geographic health measures applicable not just to a single provider but across an entire community.

Of course, even as organizations make and execute plans to cull and analyze data, our understanding of what risk factors matter and what treatments work—or don’t —is always evolving. Here, too, looking at EMRs across populations opens up new opportunities.

Remember the Vioxx debacle? Weiner asks. One of the first indications that Merck’s blockbuster arthritis medication might be linked to heart problems came from EMRs.

Kaiser-Permanente and the FDA were working on a joint study of 1.39 million Kaiser enrollees when they discovered that the risk of serious heart disease for patients taking high doses of Vioxx was more than triple that of patients taking a rival drug. Kaiser has since joined forces with dozens of other health care organizations to create a “virtual data warehouse” for multi-institutional research projects.

It’s this kind of success story that Weiner loves to share. It highlights the inherent overlaps between the aims of public and private health providers, and it suggests why, among all the data sources that he holds dear, EMRs rest closest to his heart.

“[They] will one day capture everything that’s known to the medical world,” he says. And pulling out key data to analyze and act on—which can take years today—will happen nearly instantaneously, at almost no extra cost. “Making sure that all this helps to improve the public’s health,” Weiner adds, “That’s our vision.”


  • Ed Childs

    Nashville T n 06/12/2013 01:09:57 PM

    I have worked for 8 years trying to address the needs of patients with multiple chronic conditions. Using ICD and CPT diagnostic codes matched with prescriptions and then scanned with episode-based predictive software works best. Long-term solutions require individual medical records monitored by nurse navigators.

* = required field

Read about our policy on comments to magazine articles.

design element
Online Extras
Lessons From the Master

Lessons From the Master

Patti Gravitt reflects on the science and life lessons she learned from Keerti Shah during the decades they’ve spent investigating human papillomavirus.

Watch Now

Talk to Us

Amazed? Enthralled? Disappointed? We want to hear from you. Share your thoughts on articles and your ideas for new stories:

Download the PDF

Get a copy of all Feature articles in PDF format. Read stories offline, optimized for printing.

Download Now (3.3MB)