We live in a connected world. When trying to determine the influence of health systems and physicians, simple analysis based on claims or prescription volumes isn’t enough.

For instance, some physicians are high volume prescribers but don’t really influence their peers. Conversely, some doctors, who are highly connected, may write a small number of prescriptions themselves but influence many more. Similarly, when targeting rare diseases, mapping patient journeys across providers is important for effective interventions. Relational databases aren’t built to easily capture the connected nature of entities easily.

With a graph database accessing nodes and relationships is efficient and seamless. By allowing users to quickly traverse millions of connections per second, graph databases can help commercial life-science teams derive deeper and more targeted insights.

Real world healthcare data on graphs

Real world events in healthcare can be mapped to create unique relationships between physicians, healthcare organizations and vendors. This gives us a unique perspective into how patient care is handled, money flows, and organization hierarchies. Healthgraph uses public datasets like published papers, NPPES/CMS datasets, regulatory filings, websites and private datasets including claims to generate this unique graph of relationships between each entity.

Each network in isolation is easy to map on traditional databases and gives you gives a limited view. Below are examples of some controlled networks that are generally used for analysis.

Referral networks

Referral networks allow you to see the flow of patients between physicians. But in reality, this is a manifestation of other social and professional relationships that physicians may share.

Research networks

Publication-citations and co-authorship networks capture connections via research activities. This can be a key dimension of professional visibility used to research on academic and research milestones done by practitioners.

Clinical networks

The organizational structure that links a physician to a facility or a health system is an important lever in prescribing behavior.

But the real insights are derived with a complete view of relationships between HCPs and HCOs. For instance, Healthgraph combines all the above networks into a single primary directed graph for all the analysis. This graph currently contains over 34M nodes and 200M relationships. Each relationship is scored based on locally normalized values.

An application: mapping influencers through graph analysis

Finding key opinion leaders (KoLs) who are influential in particular diagnoses areas is important for pharma brand strategy. Correctly identifying the physicians who diagnosis the right disease state and who have a large influence sphere, can help life science teams optimize their outreach efforts.

Conventional approaches have relied on identifying these KoLs based on decile segmentations, i.e. the volume of claims or prescriptions that they write. But influence isn’t based solely on claims. It also depends on the network reach of a physician based on their networks:

  • Research networks: clinical trials and publications
  • Professional networks: clinical affiliations, co-workers and referral
  • Social networks: committee memberships, online networks, medical school or residency
  • Company networks: speaking engagements for companies, industry bodies

With a graph-based dataset like Healthgraph, it’s now possible to map the real-time network of practitioners across all these connections, to create an influence map that computes the true reach of a prescriber.

An example of KoL mapping of hematologist

The image to the left shows a heterogeneous network of organizational relationships and events based relationships are distilled down to the HCP<->HCP relationships using all available paths between each HCP

Each of the relationships or combination of relationships are scored (can be tailored to the customer needs) based on the type and weight of the individual relationship. This network then runs through a combination of scoring parameters including Eigenvector scoring for nodes and influence scoring for nodes and relationships.

The post is part 2 from a series of articles about Graph-based data, databases, and algorithms

Previous – Part 1: A brief history of graphs

Next – Part 3: Scoring nodes and relationships in a healthcare graph

Want to know more on how Compile can help with your commercial strategy? Talk to us.