Predictive analytics is about finding patterns, riding a surfboard in a data tsunami

Artificial intelligence and machine learning are fast accelerating the insights that can be gleaned from the granular descriptions enabled by ICD-10.
By John Andrews
10:32 AM
Share

That dense pile of unstructured data may look like an impenetrable universe, but upon closer examination, the mammoth mass of raw information holds insights that can greatly benefit healthcare. It comes down to identifying logical patterns within the chaos and extracting them for analysis, experts say.

As data analytics progresses, researchers are learning more about how to harness the massive amounts of information being collected in the provider and payer realms and channel it into a useful purpose for predictive modeling and population health management as well as for a multitude of clinical and administrative functions.

Paul Bradley, chief data scientist at Louisville, Kentucky-based ZirMed, jokingly refers to riding "a geek surfboard" in his job of probing healthcare's data tsunami. But his job is the equivalent of finding meaningful droplets within that tidal wave. The scope of that task is intimidating, but it is within these microscopic segments that meaningful discoveries are made.

ZirMed is working with providers in various stages of predictive analytics adoption and Bradley sees the larger health systems benefitting the most at this point.

[Also: CMIO says analytics, artificial intelligence, Big Data and machine learning are becoming IT requirements]

"They have the raw material and the EHRs collecting the information, so it comes together," he said. "Predictive analytics takes the data and finds opportunities based on patterns and trends. It finds pockets of revenue by studying the granular level."

Patient-physician encounters can have thousands of attributes available for analysis from a billing perspective, services rendered and medications prescribed. Once the data set is prepared, ZirMed applies predictive modeling through charge integrity to extract patterns or trends.

"If an orthopedic surgeon uses certain bolts or plates to repair a joint along with specific drugs, there is a profile that can be created," Bradley said.

The more granular descriptions within ICD-10 has also helped zero in on missed revenue opportunities, he said.

"Missing just 1 percent to 2 percent of charges on claims in big health systems can come out to millions of dollars a year," Bradley said. "Across the spectrum, our clients have seen ROIs of four- or five-to-one."

Getting lingual with it
Natural language processing is another application that explores the contents of unstructured data and turns it into structured data for actionable insights. Boston-based Linguamatics focuses on "rapidly converting the valuable information hidden in text into knowledge," says Simon Beaulah, senior director of healthcare.

The unstructured text of EHRs and literature in the life science domain are also sources of extraction, with Beaulah using ejection fraction as an example.

"If ejection fraction is less than 45 percent in the left ventricle of a poorly performing heart, that is a really important measurable value that indicates a serious disorder and is in the clinical notes," he said. "The reference to the ejection fraction needs to be identified – EF or Multiple EF. Once found, it can be extracted and turned into a discreet data element and applied to the overall congestive heart failure population to determine the appropriate treatment. The same can be done for pulmonary function in COPD or cancer staging and is very important in determining outcomes."

Without this technology, the process requires a manual chart review and reading through reams of notes to track down the "trapped information," Beaulah said.

Welcome to the machine
With all the attention artificial intelligence/machine learning is getting, Gurjeet Singh cautions that the public "underestimates where it is and overestimates where it will be."

What it will be, says Singh, CEO of Palo Alto, California-based Ayasdi, is a technology that accelerates data cultivation without human interference. (Where it won't be, he adds, is the worst-case scenario of machine self-awareness in The Terminator film series.)

Already in use by the world's biggest banks, Ayasdi's machine intelligence platform ingests and processes large volumes of internal or third party data, then applies multiple machine learning, statistical and geometric algorithms to gain insight and predict the future. It got its start at Stanford University with the Human Genome Project, which through a $10 billion investment mapped out the first seven genomes in detail.

"The challenge started with how to handle large data sets," Singh said. "We didn't know what we'd find. What we realized was that people weren't asking the right questions in a complex data set."

Patient-provider interactions are complex data sets, but for defragging the massive cyber-lump "analytics aren't that useful," Singh said. "It's about building apps that go to the dashboard. The problem with analytics is that the people who do it are not the ones who use it on a day-to-day basis. We need to build apps for the professionals who can use it themselves."

In explaining machine intelligence's role in healthcare, Singh refers to four "notions": Discovery, prediction, justification and action. "Any intelligence system that embodies these four concepts will use data efficiently," he said.

[Also: Healthcare data: A beast best tamed by machine learning?]

The artificial intelligence advantage is the removal of human subjectivity from the equation, though machine logic alone is not enough – "it needs to go hand in hand with human guidance as well," Singh said.

Zip code mining
Quantifying healthcare has one thing in common with real estate: Location, location, location, says Dave Hom, chief evangelist for SCIO Health Analytics in West Hartford, Conn. Specifically, he says it comes down to the zip code as a fertile source of demographic information about patients.

"The zip code is the greatest predictor of health status than genetics or education," Hom said. "Zip codes show income levels for towns that have vulnerability to diabetes. Zip codes with lower incomes typically have lower education. The number of physicians is much lower in low income areas and the number of fast food places is higher. Healthcare needs to realize the value of the zip code."

Zip codes are rich sources of information about their residents – income level, education level, number of dependents, spending habits and other factors that can be critical to understanding a patient's ability to comply with physician orders, Hom said, adding that other social determinants like unemployment, crime, even weather patterns can impact behavior. Moreover, he said the four-digit zip code extensions can create even more finely detailed micro-profiles.

Understanding the zip code and how to layer it onto the data creates value in measuring patient risk and propensity to consume," Hom said. "For physicians being paid for outcomes, getting that data helps them understand the viability of the patient."

Just getting started
The analytics movement in healthcare is "at the beginning of the beginning" and Singh says he is "super excited about its potential…there are some things we can attack right away, but there is so much more. Hospitals have barely started collecting genetic information and it will become super impactful once we learn more."

For Beaulah, the future represents "more opportunities to improve care," pointing to behavioral health and its co-morbidities as an unexplored frontier of unstructured data.

Ultimately, analytics needs to reach a point where "systems are built for physicians to give care their way without the need to explain it," Bradley added. "It's a tall order, but the forces out there around the escalating costs of care and a boomer population that increasingly needs it will enable the right technology to be developed."

Twitter: @HealthITNews


Like Healthcare IT News on Facebook and LinkedIn