Before approaching high-level predictive or prescriptive analytics, successful healthcare organizations must lay the right groundwork. However, many organizations struggle with how to do so.
The Parkland Center for Clinical Innovation, a health IT research and development think tank created by Dallas-based Parkland Health & Hospital System, has experienced this challenge first-hand. PCCI operates as an separate entity, employing its own data scientists to develop artificial intelligence and predictive analytic models for disease detection and personalized medicine, among other initiatives.
"It's definitely a really exciting time, from the data analytics perspective, to be in healthcare," says PCCI CEO and President Steve Miff, PhD. "But a really close working relationship between the data scientists and the clinicians is critical."
Four members of the PCCI team spoke with Becker's Hospital Review to discuss five data challenges they wish their healthcare partners understood.
Here's what they had to share.
1. Data is a spectrum. Industry experts, including Dr. Miff, say actionable healthcare data runs on a continuum in which data is able to answer different questions.
The spectrum includes descriptive analytics (which answer the question: What happened?), diagnostic analytics (which answer the question: Why did it happen?), predictive analytics (which answer the question: What will happen?) and prescriptive analytics (which answer the question: How can we reach the optimal outcome?), according to Gartner.
"Most of the industry has been focused on descriptive analytics and diagnostic analytics," Dr. Miff explains. "We've been starting to see the move toward predictive analytics, but from an industrywide perspective, we've barely passed diagnostic analytics."
2. Data needs to be integrated. When developing predictive models, PCCI often brings together data from community organizations, insurance claims and EMRs. However, lack of interoperability hinders researchers' efforts to create accurate and complete datasets.
"There's not a national patient identifier, so it can be very difficult at times to link up patients across the continuum at all these different touch points," says Medical Director Vibin Roy, MD, who works with the data science team on projects related to patient quality and safety. "You might notice issues with data quality or have trouble identifying a patient."
Without quality information, it is impossible for a data project to get off the ground. Albert Karam, an associate consultant and data scientist who develops predictive models for Parkland Health & Hospital System and its associated health plan, stresses a similar point: At the end of the day, "if you don't actually have the data, or you don't have quality data going back far enough, then you can't be effective."
3. Data doesn't come ready to use. Even after a research team aggregates a complete dataset, the majority of its time isn't spent on predictive modeling. It's spent preparing the data for analysis by putting it into an appropriate format.
"Data is only as good as how it's collected," says Shelley Chang, MD, PhD, a data scientist and physician. "At the very beginning, when you get the raw data, you need to do a lot of work to validate that the data is telling you what you think it's telling you."
Dr. Chang says her first step is learning how a source collects data. Two hospitals, for example, might use two different clinical terms to describe the same condition. Or, one hospital might have changed the way it records patient data over time by adding or consolidating codes in its EMR.
In addition to reviewing existing data, preparation time also involves working with information that isn't visible.
"One of the things that's really time consuming is deciding what to do with missing values," Mr. Karam says. "If we're looking at temperature data, and for one person it has no data, or says something like 0 degrees Fahrenheit, how do we deal with that? It's something that takes a lot of consideration and a lot of thought, because if it's done wrong, then we're changing the data in a way that's not fair."
4. Data needs robust infrastructure and systems. When working with large quantities of information, data scientists are slowed by technology that lacks the capability to handle "big data."
Sometimes, these challenges arise from the hardware itself. "I have a couple projects where I have 12 million rows of data, or 26 million rows of data, and my computer can only handle so much," Mr. Karam says. "The process ends up taking four hours. You need to have architecture and engineering in place to maintain and handle these millions of rows of data."
Other times, it means data scientists need to spend additional time developing models to accommodate changing data systems, such as the recent change from ICD-9 to ICD-10 codes. A hospital needs to keep data scientists abreast of changes to its data collection practices so researchers can adjust their model accordingly.
Dr. Chang highlights how data scientists must create conservative models, which are "robust enough to handle these changes in data sources." "Otherwise, you could potentially have predictions that are not accurate," she says.
5. Data scientists and hospital leaders need to work together. A researcher can create a predictive or prescriptive model alone. However, for the model to be useful, data scientists and hospital leaders must collaborate to ensure it can be deployed at the point-of-care.
"When we develop a model, we can certainly come up with predictions," Dr. Roy says. "But the question is: How useful is that prediction in a hospital's day-to-day workflow? If we better understand their issues and their workflow, we can really try to tailor the solution to what they're looking for."
For Mr. Karam, understanding a hospital's workflow and where the predictive model will fit in is integral. "Going in with just a pure data science perspective isn't enough," he says. "You can sit there and play with data all day, but it's not going to mean anything until you actually know how and where it's going to be used."