Every healthcare stakeholder agrees that it’s time for clinical data to get up and running. But just because everyone’s behind an effort certainly doesn’t make it easier.
Clinical analytics, at this stage in its evolution, is really a race against time. For provider organizations, the future of value-based care came much faster than expected: Medicare recently announced an ambitious timeline for transitioning to value- and risk-based reimbursement models, and state governments and private insurers have been relentless in their efforts to tie reimbursement to quality, with the overriding goal of cutting their total costs of healthcare.
In response, providers are scrambling to embed clinical analytics into their networks to identify current and future high-risk patients. With so much revenue at risk, letting a handful of patients slip through the cracks and rack up huge treatment costs can sink the best efforts to control the costs of large populations.
For example, at Atrius Health, a push further into population health has unveiled a “gray area” within its patient population that typical risk-modeling and clinical analytic efforts overlook, says Joe Kimura, MD, chief medical officer at the Boston-based not-for-profit alliance of medical groups that comprises 42 locations, 750 physicians and 6,800 employees.
“With the tools available now, it’s pretty simple to know your highest-risk patients, because they’re in and out of your facilities and the hospitals—you have a lot of encounter data to work with,” he says. “But we’re pretty much at full risk for our patient population, so we need to get into that gray area of our population that we don’t have much of a dialogue with, typically males between 20 and 40 years old who almost never come to see us, and when they do, they come in for something minor, like a sprained ankle.”
Joe Kimura, MD, (right) chief medical officer at Atrius Health, discusses findings with Michael Macina, MS, the system's behavioral health specialty administrator.Atrius Health is using natural language processing software (Kimura declined to identify the vendor) to analyze unstructured progress notes and other unstructured text within the EHR to extract clinical concepts to run through its risk algorithms. “Even when someone comes in for a minor event, we do what we always do, and complete a note, and within that note is information about lifestyles and socioeconomic issues and other clues about whether the patients are at risk for diabetes and obesity,” Kimura says. “There’s so much information is in that narrative that can be used by our algorithms for patients that in the past wouldn’t be flagged because we didn’t have enough events to key a picture of them.”
Many healthcare organizations are finding analytics efforts difficult to undertake.
At University of Mississippi Medical Center, executives were behind a big push to move into population health management and predictive analytics to respond to the incessant financial pressure to lower costs while keeping patients healthier. So the medical center started assembling the data sets deemed necessary to analyze its population.
The effort to build up data sets took seven long months. When the data was compiled, the organization began the process of scrubbing and eliminating outliers, filling in null values and otherwise transforming that data in a number of ways. For John Showalter, MD, UMMC’s chief health information officer, the process seemed to be taking forever.
After talking through the problem with Jvion, a predictive analytics firm with which the medical center worked, Showalter took UMMC’s effort in a dramatically different direction—instead of spending months cleaning up the EHR data, UMMC started sending Atlanta-based Jvion raw “dumps” from its EHR, which Jvion then feeds into its machine-learning network and adds numerous other clinical and socioeconomic variables, such as U.S. census data and information from credit reporting firms.
What comes out are predictive risk scores for a range of health issues UMMC is targeting, including heart disease, hospital readmission, pressure ulcers, blood clots and hospital-acquired infections. The risk scores are brought back into the medical center’s electronic health record, from Verona, Wis.-based Epic Systems, and are visible to physicians when they’re in the patient’s record. At that point, the physician can put a patient on a risk protocol for a certain condition, and trigger orders and alerts for other caregivers.
Quote“I think everyone is trying to figure out how to reduce the ‘time to value’ of their data."
“Our overriding goal was to get advanced clinical analytics up and running, and when I started seeing what could be done with the raw data we were sending, I realized that initially we were going about it in the wrong way,” Showalter says. “We’re at the point now where we can tell Jvion we want a new use case with our data, and in six to 10 weeks we can have a new prediction built into our infrastructure.
“I think everyone is trying to figure out how to reduce the ‘time to value’ of their data. The data’s not perfect, and it never will be, but with the machine-learning tools and neural networks out there today, it doesn’t have to be perfect to be meaningful. For example, you might have a patient whose blood pressure is recoded as being 1,000, which is impossible, but machine-learning can be taught to ignore that and weigh the other data points about that patient for risk. So if it’s looking at 10 data points to create clusters of patients at risk, it will forget about the blood pressure and determine that nine other pieces of information about that patient put them in a specific risk cluster.”
Many providers have found their clinical analytic strategies are hitting a wall because they’re not well-connected to the true agents of change in medicine—the caregivers who are now living their days in EHRs and other information systems, all the while being deluged with data and documentation. That disconnect has had a cascade effect on the development of clinical analytics, as providers, unhappy with the value yielded by their analytic efforts, are starting to rethink their strategies.
Some are now putting more resources into building their own analytics systems by developing algorithms and tools; others are trying out new technologies to get clinical analytics closer to caregivers. The result, according to Joe Van De Graaff, research director at KLAS Enterprises, is a very distinctive shift in the analytics market.
“It’s really a piranha tank right now,” he says. “We’ve spoken to several hundred health systems in the past 18 months, and we estimate that up to 30 percent of them are looking to replace their population health and analytics products. There’s a very strong need to get analytics into workflows, and many organizations are uncertain if their current analytics platforms will be able to do that.
“A few years ago, there wasn’t a big focus on using a core EHR platform for business intelligence and analytics, but that’s changed dramatically, as organizations have moved deeper into new care and reimbursement models.”
Chicago-based NorthShore University HealthSystem, for example, has been building up its stable of homegrown analytics tools for the past few years, and it’s learned through trial and error that to be effective, analytic tools have to be visible—or more to the point, visual, says Ari Robicsek, MD, the health system’s vice president of clinical analytics.
“Traditionally, what we would do is send a report out once a month to someone about how their department is operating, but that wasn’t providing analytics for the clinical workflow, which is where it really needs to be,” he says. “What we’re focusing on now is creating visualization dashboards that let them explore the data themselves, and integrating those visualizations right into their workflow.”
NorthShore—comprising four hospitals and 2,100 affiliated physicians, including a 900-physician medical group—recently rolled out a number of applications that are crunching massive, real-time data sets to come up with predictive analyses that enable them to focus attention on patients at immediate and longer-term risk for health problems.
Quote"You’re seeing some health systems with the resources and know-how deciding they can build clinical tools better than anyone else in the market."
According to John Moore, founder and managing partner at Boston-based healthcare analyst firm Chilmark Research, there is a dearth of vendor products on the market designed for clinical analytics, and few if any of those are integrated with electronic health records and workflow.
“There really aren’t a lot of good options out there—EHR vendors are starting to build their own clinical analytics into their products, but that’s an effort that’s really just getting started,” Moore says. “For that reason, you’re seeing some health systems with the resources and know-how deciding they can build clinical tools better than anyone else in the market. It’s a reflection of the maturity of the marketplace for analytic products that utilize clinical data—mostly what’s being offered are standalone products that don’t mesh with workflows.”
NorthShore uses visualization software from Seattle-based Tableau—the software sits atop its enterprise data warehouse. The health system has developed a number of predictive algorithms and analytic tools, and linked them to its Epic electronics health records system to enable case managers and physicians to access multilayered visualizations and utilize that information on the go.
For case managers, a predictive algorithm on a daily basis analyzes data flowing into the EHR about hundreds of thousands of patients, and provides a dashboard look at the patients at highest risk for, as Robicsek puts it, “bad things happening to them,” be it a hospital readmission, acute psychiatric problem or cardiac issue. From there, case managers can drill down to understand why the patients are deemed at highest risk and decide whether they should be enrolled in a case management program. The predictive tool also helps case managers, based on all available data, determine what day and time of day is best for making that initial contact with the patient. “We’ve found that it’s extremely important that case managers have an idea what the optimal time is to reach out to a patient and make that first connection,” Robicsek says. Once a patient is enrolled in a program, the visualization tool tracks his or her progress or lack thereof by highlighting clinically significant changes in health status.
NorthShore has also gained a lot of traction with clinical analytics via links with its EHR while utilizing Tableau’s visualization software. It recently rolled out an application called the My Panel dashboard that enables physicians to click on a button within the EHR that takes them to a visualization of how they are performing on a number of different clinical quality metrics for their patients. They can also move to a view that displays the predictive model data for each of their patients and what the risk score is for readmission within 30 days, as well as other potential risks.
Color bars show whether they are above or below certain targets, such as how their patients are controlling their diabetes or hypertension. They can then create lists of patients who are not on target for each metric, and hover the cursor over each individual metric to display care gaps and guidelines for each. Via another button, a physician can send a direct message to an individual patient to tell the patient to schedule an appointment, or send messages to other caregivers to order tests and other services. Within that same dashboard, they can view a map that shows them where their patients live and what types of health services are nearby.
The dashboards do not actually reside within the EHR; clicking the My Panel dashboard takes users to a web page that runs the visualizations. But the experience is seamless to physicians, Robicsek says, and links embedded in the visualizations take users right back into sections of the EHR. “Since those visualizations and predictive tools live within the skin of the EHR, physicians feel like they’re working in the same place, and there are no additional steps in their workflow. Since we rolled this out, 100 percent of our physicians have used the My Panel dashboard, and 75 percent are repeat users. In the world of physician technology adoption, anyone will tell you that’s absolutely huge—and it shows that no matter how good a predictive algorithm or other clinical analytical tool is, it won’t get adopted unless it’s right there in that clinical workflow.”
Getting analytics embedded into EHRs and subsequently the workflow of providers is a key element for many health systems, but there’s a two-way street there: Within the EHR itself is a trove of unstructured data that can be utilized by ever-more sophisticated analytics engines to assess the patients’ health risks.
Natural language processing is also a key technology in the University of Mississippi Medical Center’s effort to widen the analytics net, says Showalter. “Getting the clinical concepts out of that data is really key at the population health level,” he says. “It flags a lot of the problems that we can’t really identify with our structured data. For example, if someone comes in for abdomen pain, the scan might find that it’s being caused by a kidney stone, but besides the stone there is a 6-centimeter dilation of the aorta, which has to be surgically repaired. That’s the type of information that isn’t put into a field but is clinically important.”
Showalter says the NLP platform, from Franklin, Tenn.-based M*Modal, can analyze about 98 percent of the medical center’s unstructured EHR data. “The only data we can’t really bring into our clinical analytics effort are waveforms. We don’t have great use cases on how to handle that information, so I guess that will be our next hurdle to get over.”
(This article appears courtesy of our sister publication, Health Data Management)