UPMC turns to NLP to make sense of unstructured data
Faced with the reality that 80 percent of healthcare data is unstructured, the University of Pittsburgh Medical Center is investing in natural language processing to leverage this untapped resource.
UPMC Enterprises, the venture arm of the Pittsburgh-based health system, is focused on solving problems in healthcare and driving innovation by harnessing big data.
The organization sees NLP technology as having the potential for making the abundance of unstructured data in healthcare accessible and actionable.
At the epicenter of NLP activities at UPMC Enterprises is Rebecca Jacobson, MD, vice president of analytics, who spent 16 years as a professor of biomedical informatics at the University of Pittsburgh before joining UPMC.
“There are different areas of NLP, but the one we primarily work in is information extraction,” says Jacobson. “At least three of our portfolio companies that we work with closely are using natural language processing, including medCPU, which is a company that is doing real-time clinical decision support using NLP.”
In 2016, UPMC took majority ownership in New York-based medCPU, whose Context Engine is touted by the vendor as solving the “traditional NLP accuracy limitation” by providing “95-plus percent accuracy of content extracted from free text—recovering 30 percent to 90 percent of information not retrievable by other systems.”
In addition, according to the company, its medCPU Reader “accurately and continuously acquires all information including free-text notes, dictations, and structured documentation from electronic health records and ancillary systems without the need for high-level interfaces or integration on the client’s part.”
“They have an innovative model where they are able to see what is happening on a clinician’s terminal as they are typing and ordering things—without directly integrating into the EHR,” adds Jacobson, who notes that medCPU’s clinical decision support tools take the data entered by clinicians in real time, analyzes it and provides point-of-care alerts, as well as shares best practice recommendations.
San Mateo, Calif.-based Health Fidelity, another of UPMC Enterprises’ portfolio companies, is focused on clinical NLP technology.
“That is a company that’s doing risk adjustment for healthcare payers,” comments Jacobson. “Specifically, they use NLP to make (Hierarchical Condition Categories) coding much more accurate.”
According to Health Fidelity, it offers the “only solution on the market that utilizes NLP to extract valuable insights from medical charts, changing the way risk is identified, quantified and managed” at health plans.
Before UPMC became an investor in Health Fidelity, UPMC’s Health Plan tested the NLP approach to optimizing risk adjustment on the payer side and in the process increased coder productivity and added about $200 million in revenue.
Last year, UPMC had selected Health Fidelity’s HF360 Provider, which it contends is the industry’s first NLP-powered, EHR-integrated physician workflow solution for risk and quality gap closure. UPMC is using HF360 to standardize the risk capture process across all populations and increase the efficiency of that process through targeted patient interventions.
When it comes to technology that has not yet been commercialized, Jacobson points to an ongoing project between UPMC Enterprises and UPMC’s Wolff Center designed to create a new NLP-based system for rapidly abstracting clinical quality metrics.
“There are so many different quality programs on the payer and provider sides that are difficult to compute, are costly and not so accurate,” contends Jacobson. “You have to do them manually using human abstractors. But, over the last year and a half, we’ve been developing technology to be able to do that abstraction in a partially automated way that makes it much faster. Now, we’re looking to take some parts of the process and make it automatic so we can see things in the data earlier.”
Another area that Jacobson sees as potentially benefitting from NLP is the identification of social determinants of health within EHRs.
“It’s been argued that 60 percent to 70 percent of health outcomes are driven by factors such as your social support network, access to stable housing, food and transportation, and the degree to which you are exposed to violence or substance abuse,” she says. “But this information is hardly ever captured by EHRs within explicitly structured fields. Identifying and extracting this information from the clinical notes could help to improve predictive analytics and eventually help us to intervene more effectively.”