Novo Nordisk: Natural Language Processing On A Steady Diet Of Real-World Data

By Deborah Borfitz

July 29, 2019 | Natural language processing (NLP) is finding novel applications at Novo Nordisk, turning a trio of real-world data (RWD) sources into a goldmine of knowledge about the medicines it makes, the healthcare providers who prescribe them, and patients who usually pay for a portion of the cost. Novo Nordisk leverages the NLP text-mining tool of Linguamatics to create efficiencies and generate actionable insights enterprise-wide, winning the company an Innovative Practices Award at the recent Bio-IT World Conference & Expo '19 in Boston.

NLP is most typically used within the life sciences industry to select targets for the drug discovery process, according to Thierry Breyette, associate director of information analytics at Novo Nordisk. Some companies have also begun using NLP to assist with regulatory review and compliance activities, and Linguamatics has published case studies describing the use of text analytics to better understand adverse events, and design and interpret studies.

The use case at Novo Nordisk may be the first time that the text being mined comes from real-life conversations—specifically those its medical information and liaison teams have had with healthcare professionals, and interactions between customer call center staff and inbound callers over drug-related questions and concerns, says Breyette. What has been gleaned from medical information requests and field medical affairs notes has been particularly illuminating for both in-house medical and commercial teams because the providers being engaged are typically key opinion leaders in a community, he adds.

NLP can quickly spot a product-related issue or question that is trending—e.g., how to titrate a medication or switch a patient from one medication to another—so the medical information team can start creating content providing answers and medical liaisons can better prepare for upcoming consultations, says Breyette.

Product teams can use NLP to learn what questions providers were asking during an initial product launch to decide how to engage them in later stages of the rollout, Breyette says, as well as ensure field staff is trained and prepared for those conversations. The publications team also pays attention to providers' FAQs when mapping out publishing goals for the year.

Patterns detected in customer call center reports might reveal a training opportunity for the sales force, Breyette continues. If people start calling in to complain a product isn't covered under a patient assistance program, physicians may need reminding that their patients must be pre-qualified for the program.

One of the biggest surprises to date, Breyette says, is the degree of regional variation. The way insights break out across the country can be strikingly different depending on factors that include access to health insurance and disease prevalence. And trends can be layered against larger datasets, such as those maintained by the Centers for Disease Control and Prevention, to produce a fuller picture of healthcare realities place to place.

Migration to AWS Platform

The RWD sources linguistically processed with I2E text analytics include medical information requests (20,000 per year), field medical affairs notes (3,300 per year) and customer call center reports (130,000 per year), reports Breyette. Queries related to topics such as safety, efficacy, pharmacokinetics/pharmacodynamics, randomized controlled trials, patient populations, dosing and devices were refined in close collaboration with internal subject matter experts to ensure the desired level of precision and recall.

Results are produced in a standardized, structured format and fed to Tableau to create self-serve, drill-down dashboards, explains Breyette. These "medical patient dashboards," as they're called internally, are in routine use by more than 50 Novo Nordisk users worldwide sharing the insights with a much larger universe of data analytics consumers—including Diabetes Care and Education, market development, and commercial effectiveness teams. The data visualization software might be used to not only learn what safety- or efficacy-related questions healthcare professionals are asking about a drug, but if the same questions are being asked about competitor products.

Migration of Linguamatics I2E and Tableau to Novo Nordisk's cloud-based data and analytics platform OASIS (open analytics & insights) was a key enabler of seamless data-sharing, says Breyette. OASIS, built on Amazon Web Services (AWS) and managed in collaboration with services from Accenture, is focused on making data readily available from essential repositories.

Automation was virtually impossible when applications were sitting on multiple incompatible systems in different parts of the organization and "everyone owned a piece of it," says Breyette. Information was scattered across the company, forcing people to manually sift through documents and create static "Monthly Insights" reports.

It was tedious work that was often outsourced, a significant expense equivalent to the cost of a full-time employee, says Breyette. Moreover, this manual curation was not consistently done by the same people, and what constituted an insight was not standardized. Insights collected from call center interactions also weren’t being broadly shared across medical and commercial teams.

By moving to a cloud-based environment, software tools are available in one spot and data can amass in a common lake, says Breyette, plus Novo Nordisk doesn't have to own all the servers and software. Using AWS, the company has access to a multitude of tools it may potentially need—and not necessarily through a long-term licensing commitment. More data can be easily onboarded as demand for the dashboard ramps up among other people in the organization, he adds. "Everything is in a template now. It's just a matter of plugging in the data… it's not a new project every single time."

Novo Nordisk clinical trial management system also serves as a data source for the NLP tool, to help improve the quality of studies, Breyette says. Linguamatics I2E can use free-text site visit data entered by monitors to efficiently classify protocol deviations on a more granular level, informing site-specific mitigation plans. If an informed consent deviation has occurred, for example, trial managers know "exactly what went wrong… [for example], a signature process or the ICF version." That information might also feed back into decisions about site or investigator selection, he says.

In addition to significantly broadening access to on-demand insights to enable evidence-based decision-making across a global team, Novo Nordisk has reduced the burden of manually scanning documents so people can focus on "higher value" work such as actual analysis and communication of trends, says Breyette. Vendor spend on report creation has also fallen sharply.