Provider organizations are using computing power to determine which patients need help—before crises occur.
Analytics allow providers to identify— and tighten—gaps in care for at-risk patients.
Anyone who’s shopped on the ecommerce juggernaut Amazon knows that the company analyzes customers’ shopping habits and serves up suggestions for what they might like to buy next. If you’re looking at one type of product, the site can tell you how many customers bought a different item or what they bought in addition to that item.
“Why can’t we do that in healthcare?” wonders Charles “Chuck” Christian, vice president of technology and engagement at the Indiana Health Information Exchange (IHIE).
Organizations are “learning now how to standardize [data] and normalize it using good standards and being able to use that data to improve the quality of care and the quality of the outcomes,” Christian says.
The next step is to serve it up to physicians before they even need it to get patients the right care at the right time and to use predictive analytics to keep populations of patients healthy, he contends.
If we can predict shopping preferences, we should be able to put that technology to work to predict when someone is at risk for readmission or might miss an appointment, adds Deborah Viola, vice president of data management and analytics at Westchester Medical Center Health Network, a nine-hospital system based in Valhalla, N.Y.
“That’s what makes this work exciting, because it has so much potential,” she says.
Around the country, healthcare organizations are taking that next
step—using data to predict, intervene and prevent health crises. The goal of such initiatives is to serve at-risk patients across the care continuum by reducing or eliminating gaps in care, to prevent readmissions or costly and unnecessary emergency department visits, and to improve population health with a focus on prevention for those who are most in need of targeted intervention.
Machine learning, artificial intelligence and predictive analytics are helping healthcare in this quest.
Opening the toolkit
Geisinger Health System, for example, uses a variety of tools to power its data analytics program. The 13-hospital integrated system, based in Danville, Pa., has been analyzing historical records gathered over more than a decade using machine learning and artificial intelligence.
Geisinger uses two big data platforms: the open-source Apache Hadoop and Cerner’s HealtheIntent platform. About 87 percent of its data is not discrete—it must be mined from the text in sources such as physician notes.
“We use natural language processing to scan through those notes, looking for specific elements that we can extract and use for analytics purposes,” says CIO John Kravitz.
Radiologists’ notes, for example, are rich with data about imaging results, such as the size of the aorta and deviations of the aortic wall. “We look at those notes again and compile that into a database. And if we see, for any reason, that the nodule is growing in size, that triggers events from predictive analytics to identify to the surgeons that this patient will most likely have an aortic aneurysm in the near future,” Kravitz says.
When it was first rolled out, that program alone saved the lives of 27 people, he says.
Atrius Health also uses natural language processing to create queries and extract clinical data from free-text fields, using the Linguamatics platform. For example, it queries unstructured echo reports to analyze heart pump function to identify patients with a high risk of heart failure.
The initiative has enabled the organization to better identify gaps in care, says Craig Monsen, CMIO of the nonprofit organization, which comprises 32 physician practices with more than 800 doctors across eastern Massachusetts.
The organization has documented real-world results. In 2017, Atrius
Health identified 92 otherwise undocumented congestive heart failure and chronic obstructive pulmonary disease (COPD) patients and gained more than $75,000 in additional annual risk-adjusted revenue per disease area by intervening.
Nurses also benefited with a significant increase in the efficiency of their chart reconciliation process, Monsen says. Problem lists are more accurate and complete, and ACO reporting is simplified.
Closing gaps in care
In addition to identifying gaps in care, predictive analytics offers the opportunity to identify potentially “closable care gaps,” Monsen says. From there, providers can reach out to patients, such as those who are likely to miss appointments, put off tests and procedures or fail to follow discharge plans and medication regimens, and create a care plan they can engage with, he says.
Geisinger is also working to close care gaps, especially for patients with chronic conditions, such as COPD or diabetes, using its big data platforms. It uses the data to create electronic health record order sets so that the next time the patient presents for care at any point in the care continuum, clinicians can use those sets to take action to narrow or close those gaps in care.
HIE data identifies patients who get care outside of the Geisinger system to avoid creating unnecessary order sets.
Predictive analytics offers the opportunity to identify potentially closeable caps in care.
There’s a similar effort underway in Indiana—a patient-centered data hub alerts IHIE members if one of their patients shows up in a medical location outside of the state. The exchange can request that clinical data and serve it up to physicians.
“We’re gathering data from the full landscape of healthcare,” Christian says.
Predicting readmission risk
Westchester Medical Center puts data analytics to work to develop a readmission risk prediction model.
The data analytics teams at WMC and three-hospital Bon Secours Charity Health System, which is part of the WMC network, used artificial intelligence and machine learning to analyze historical data on patient discharges and 30-day inpatient readmissions to identify patients who were more likely to readmit and determine what they had in common, using data analytics firm Health Catalyst and the open-source machine learning software Healthcare.ai.
The risk models include 24 variables
that were most predictive of readmissions within 30 days, including age, type of diagnosis at first encounter, and chronic conditions, such as diabetes or COPD.
“We look for associations and relationships; we run regression models to figure out what factors might be more likely to contribute to an outcome,” WMC’s Viola says.
“So when you take this approach, it can fill those critical gaps in chronic or preventative care by focusing your resources and attention on those patients who, when they’re discharged, may have a higher readmission risk score.
“So then you can say, ‘What is it we need to do to help that individual or to ensure that individuals are not going to readmit to us, especially if that’s not the right pathway to their care?’”
The model is 17 percent more accurate than widely used readmission risk models in identifying patients at high risk and low risk for readmission within 30 days, in part because it’s based on data from WMC’s own patient population.
“The key to health systems taking this on and learning from their own patient databases is that we can really reflect the characteristics of our patient populations. We’re not relying on these generic risk model algorithms that weren’t trained on our population. So we’re more likely to capture those factors that put our folks at risk for preventable readmission,” Viola says.
“By having this risk score, you have a sense of the patient that’s more urgent, so we can look at high-risk patients and also the rising-risk patients,” Viola says.
The customized risk model reduced false positives by 30 percent and also increased true positives, ensuring that critical resources are going to the right patients, Viola says. It also saves time—previously, care team members were pulling data from multiple sources and “triangulating” it on their own. Validated data on high- and rising-risk patients saved more than 1,200 hours per year.
“After we run this through and get a better sense of what variables are significant for our population, we use our own analytics platform, and we can bring in other data, such as clinical and billing data, against subgroups of patients to ensure we’ve identified the right factors,” Viola adds. Machine learning ensures the algorithms continue to improve.
More accurate, more accessible data enables care managers to follow up with at-risk patients faster—usually within seven days—to connect patients with appointments and services to help that will help prevent unnecessary emergency department visits and readmissions.
“If we can automate, if we can help them so that the hours they have are with direct patient care, that’s what we should be doing,” Viola says.
Leaders who are pondering how to duplicate this kind of effort should “think about how they can use the technology … and what influence technology can have in advancing the work we need to do in health systems,” she adds.
Working hand in hand with the clinical team is also important. “We just help build [the models], but the actual validation and operationalizing and information is really done by those teams,” she says. “This is such a collaborative process. That’s what really makes this work.”
Predictive models like WMC’s can also pull in data about determinants of health, from demographic and socioeconomic information to patient-generated data—and identify patients who are potentially susceptible to anything from heart failure to missed appointments.
“There are a lot of populations at greater risk, in part because of access issues,” Viola says. “If we can understand more about those patients, we can do more. We want to make sure patients get the care they need.”
Geisinger, meanwhile, is using its considerable data analytics power to tackle the opioid crisis in its region, reducing opioid prescribing rates by about 65 percent.
After an analysis of 942 Geisinger patients who overdosed on opioids found a steep increase in the use of acute care—especially expensive ED services—before an overdose, the organization used data from its EHR to identify providers with high opioid prescribing rates.
“In 2015 and 2016, we were pretty high in prescribing opioids, just like other health systems have been,” Kravitz contends. Issuing 30-day opioid prescriptions was “hardwired” when it comes to surgical patients—in other words, pretty much automatic. “There’s your key for addiction,” he says. “That’s scary.”
Now, a dashboard shows the data— which is visible to all providers in the system—including each physician’s panel size, the number and percentage of patients who are prescribed opioids, open orders the patient hasn’t filled and the daily morphine equivalent dose, for example. While an oncolo-
More accurate, more accessible data enables care managers to follow up with at-risk patients faster.
gist’s panel might include about 7 percent of patients on opioids, the goals for others, including primary care physicians, are much lower—from 1 percent to 2 percent. The system flags outliers.
From there, the organization intensified its focus on nonopioid pain treatments for patients, launching awareness programs for physicians and patients alike. For example, surgeons or surgical staff now explain the risks and alternative pain treatments to patients and their families before surgery. If a patient insists, they get a three-day prescription. An enterprise e-prescribing system helped to reduce forged or stolen prescriptions and saved a significant amount of time, money and effort, he adds.
“I’m not saying we will never prescribe opioids again—we don’t restrict prescriptions for stage 4 cancer patients, for example. But we do want to reduce it,” Kravitz says.
Predictive analytics’ future
Geisinger and other healthcare organization see the future of predictive analytics in genomic data. It could, for example, help identify patients who have a propensity for opioid addiction, Kravitz says.
“The industry is trying to identify those patients who have a proclivity for becoming addicted before they get addicted,” adds IHIE’s Christian.
In Nevada, a large population health study aims to combine clinical, socioeconomic, environmental and genetic screening data, using analytics to identify correlations that impact health, from opioid addiction to cardiovascular disease.
The Healthy Nevada program is run by the Renown Institute for Health Innovation, a collaboration between the integrated Renown Health network based in Reno, Nev., and the nonprofit environmental arm of the Nevada System of Higher Education, the Desert Research Institute, which has campuses in Reno and Las Vegas.
The organization is gathering genomic data on tens of thousands of state residents using DNA sequencing company Helix. Participants have access to the kits and their results for free, so they can take action on their own by following up with their physicians.
However, the program will also compile the deidentified data in an effort to predict who is at risk for conditions and to develop early intervention protocols. The aim is to expand precision medicine in the state—and beyond.
The pilot phase used genotyping; phase two is using exome sequencing, which provides 100 times the amount of information provided by existing genotyping or array technologies, according to Helix.
The idea is to stretch predictive analytics beyond the hospital or physician practice setting, says Renown Health CEO Tony Slonim, MD, and into the community to help patients live healthier lives using data.
“Health is a state of well-being— mentally, physically and spiritually,” he contends. “So how do you combine [environmental and other] data elements in the context of health and well-being? We are now giving people the opportunity to look ahead and not look back. And that is what drives me on this project.”
Analytics in action
Smaller organizations—especially those that don’t have the capital to spend on predictive analytics tools, staff, software and platforms—are still trying to get their heads around the business model and what services they need to keep their populations healthy, Christian says. But they can start by managing their own employees and then rolling out their findings to local employers.
If an organization doesn’t have in-house analytics resources or the money to build a team, a contract with a data analytics company could help, Kravitz adds. Other options include open-source or crowd-sourced platforms, such as Healthcare.ai or Hadoop.
“As health systems get there, these processes and this time to develop these models...will move more quickly.”
“As health systems get there, these processes and this time to develop these models … will move more quickly,” Viola says. “At first you have to just hang in there, have patience and work closely with your teams.”
Predictive analytics efforts are often aspirational, including Geisinger’s early explorations into genomics and Healthy Nevada’s goals to spread its sequencing and research nationwide. That’s good news for those organizations that don’t have the resources or expertise to take the first step. That’s because the advanced players also aspire to helping their smaller counterparts.
Scaled nationally, the Healthy Nevada program would be size-agnostic. “We’ll take on anybody,” Slonim says.
“Hopefully we’re creating breakthroughs that others can leverage and utilize,” Kravitz says of Geisinger’s efforts. “Anything we can do to help in that area, we’re just happy to help others. We’re out for the public good.” ☐