Studying Healthcare as a Network: Does the Analysis Match the Question?
Map of the U.S. depicting connections between healthcare providers who share patients and are 40 to 80 miles apart. Courtesy of Martin Zand, M.D., Ph.D.
In an effort to improve healthcare, researchers have begun to study healthcare systems like they would telecommunications or computer networks. Using network science, researchers can uncover patterns of opioid over-prescription or predict how a virus outbreak might overtax a healthcare network. However, a study from the University of Rochester Medical Center shows that the type of algorithms used to make these discoveries could provide misleading results, impacting healthcare delivery and policy.
The study, led by Martin Zand, M.D., Ph.D., co-director of the UR CTSI and professor of Nephrology and Public Health Sciences at URMC, used three different types of algorithms to analyze over 160 million Medicare Part B Outpatient claims. The study mapped how patients travel between healthcare providers, creating a “social network” picture of the Medicare system. Asking the same questions and using the same data, the three algorithms produced very different network outcomes.
Using the three different algorithms, the team built a network that linked providers or healthcare organizations by the patients they shared. The algorithms used included one similar to what is used to map the internet (trace-route algorithm), one similar to how Facebook maps friends (binning algorithm), and one that the government currently uses to construct Medicare networks.
The PLOS ONE study showed that the choice of algorithm had a huge impact on analysis, yielding very different connections between providers depending on the algorithm. Because Medicare and insurers make strategic decisions based on such analyses, the Zand’s team cautions other network scientists to choose algorithms carefully to ensure their analysis fits their research question. For example, internet-mapping-like algorithms are best suited to understand how patients flow through a healthcare network, while Facebook-type algorithms are better at identifying healthcare provider “teams” that might remain hidden using other algorithms.
Provider “teams” could reflect groups of physician and nurses working together to optimize care for a specific disease or condition, like breast cancer, or clusters of providers engaged in insurance fraud or opioid over-prescription. When analyzing healthcare teams based on shared patients, it’s important to remember that there are many factors that influence why providers share patients, including factors, like office location, insurance coverage, and unseen social connections between providers.
Zand hopes the results of this study will guide future network analysis of the healthcare system, which impacts both patients and providers by shaping healthcare delivery and policy. As of the publication of the study, the Center for Medicare Services had not released the algorithm it uses to construct its provider teaming networks, which limits research and impacts healthcare policy.
Zand and his team are now using network science to compare different state-level Medicare structures used across the US as well as provider teaming structures in different countries.
This work was supported in part by research grants from the National Center for Advancing Translational Sciences at the National Institute of Health, the Philip Templeton Foundation, and the University of Rochester Center for Health Informatics.
Read the full study here.
The project described in this article was supported by the University of Rochester CTSA award numbers UL1 TR000042, U54 TR001625, KL2 TR001999, and TL1 TR002000 from the National Center for Advancing Translational Sciences of the National Institutes of Health.
Susanne Pritchard Pallo |