Predicting Pop-Tarts: Future Applications in Radiology Data Mining
In 2004, as Hurricane Charley closed in on Florida, the CIO of Walmart, Linda Dillman, wondered which items the store should be stocking up on in advance of the storm. Employees suggested flashlights and batteries. Dillman had another idea: diving into terabytes of data on past shopping behaviors, she discovered that ahead of hurricanes, the two most-purchased items at Walmart stores were beer and strawberry Pop-Tarts. Walmart stores in Florida increased their inventories of these items, and by the time the hurricane passed over, the company had made a killing.
This anecdote was shared by Woojin Kim, MD, in Minneapolis, Minnesota, at the 2013 meeting of the AHRA. On July 28, he presented “Data Mining of Radiology Reports for Clinical Research, Education, Quality Improvement, and Operational Analyses.” Kim, who is director of the center for translational imaging informatics at the Hospital of the University of Pennsylvania in Philadelphia, told the story to illustrate the potential power of data mining—and to underscore that health care is well behind the curve.
“What is the future of data mining?” Kim asks. “Fortunately, it’s not too hard to predict. We’re 10 to 15 years behind everyone else, so we just have to look at what they’re doing today to predict what we’ll be doing in the future.”
Shortcomings in Radiology
Kim defines data mining as extracting information from a data source and translating those data into knowledge that people can use and understand. Unfortunately, medicine in general is challenged, when it comes to data mining, by the use among vendors of proprietary database formats. Radiology reports create even more difficulty, Kim says.
“The problem is that we radiologists don’t know who our patients are,” he says. “We don’t know their medical-record or accession numbers. It’s very challenging for us to search our own reports. With Google, we expect search results to come back to us in the blink of an eye, but in trying to get the same kinds of data from our health-information systems, we wait for minutes—and we think it’s normal.”
Kim also points out that many radiology applications have the kinds of cluttered user interfaces that have gone out of vogue (for good reason) in the consumer sphere. “The design of RIS interfaces is terrible,” he says. “They look like they were built in the 1980s. The generation of those who are now becoming physicians is not going to be happy with this kind of interface.”
Kim observes that tools for natural language processing, too, have a long way to go. “We should be able to mine percentage positivity of clinically important findings, laterality errors, and things like degree of uncertainty,” he says, “but the future is so much more.”
Down the road, Kim predicts, applications made possible by advanced data-mining capabilities will allow radiologists to correlate their own findings with pathology reports, improving performance while adding value. “Let’s say you read an ultrasound of the liver three months ago,” he says. “Wouldn’t you love, as a radiologist, to know whether you were right or wrong? Taking it to the next step, you can then find out who your really good and really bad radiologists are—who, in your group, is constantly calling it cancer, when it’s not.”
Experts in business intelligence from outside the health-care sphere point to self-service interaction, collaboration, predictive analytics, and complex event processing as the next areas that medicine needs to conquer on its path to making the most of big data, Kim says. Self-service interaction is empowering the average user to analyze data on his or her own, while collaboration enables users to compare their methods and conclusions using “a data-analytics tool where you can share and make comments,” Kim explains.
Predictive analytics and complex event processing will be the most powerful tools, however. Using predictive analytics (as Walmart did before Hurricane Charley), users in health care will be able to forecast outcomes for whole populations of patients. “The government is saying that it wants to reduce 30-day readmission rates, and the only way to do that is using predictive analytics. This is where health care should be going and what everyone should be interested in,” Kim says.
Complex event processing has caught on in trading, where seconds count—just as they count in health care, as Kim observes. “Complex event processing is more than a real-time dashboard,”he says. “You can use it to detect opportunities and threats right when they happen. You want to automate your responses as much as you can, and you can only do that if you can handle complex incoming data on the fly.”
Ultimately, Kim says, he envisions data-mining tools for radiology being so advanced that users don’t have to query them at all. “At night, the computers will be doing all the calculations for you, and they will be finding associations you’ve never even thought about,” he predicts.
Cat Vasko is editor of Medical Imaging Review and associate editor of Radiology Business Journal.