Most new, high-profile initiatives start with something splashy. President Biden’s proposed transformative biomedical agency, the Advanced Research Projects Agency for Health (ARPA-H), should start with something mundane: revamping the now-ubiquitous electronic health record (EHR).
Once hailed for their potential to transform U.S. health care into a modern, data-driven enterprise, electronic health records have instead increased the burden of data entry, contributed to physician burnout, and had a nebulous impact on the quality of care.
Electronic health records have become the data commons for the U.S. health care system. Every health care worker and patient interacts with these systems in some way, shape, or form. Almost every action that takes place in a hospital, doctor’s office, or other care-delivery setting is documented in them. Yet their impact has left much to be desired.
Intervention by ARPA-H could turn this around.
Modeled after the legendary Defense Advanced Research Projects Agency (DARPA), the vision for ARPA-H is to tackle fundamental problems in biology and medicine that do not quite fit existing market or funding mechanisms. Writing in Science magazine, National Institutes of Health Director Francis Collins, director of the Office of Science and Technology Policy Eric Lander, and their colleagues Tara Schwetz and Lawrence Tabak suggested this mission statement for ARPA-H: “To make pivotal investments in breakthrough technologies and broadly applicable platforms, capabilities, resources, and solutions that have the potential to transform important areas of medicine and health for the benefit of all patients and that cannot readily be accomplished through traditional research or commercial activity.”
Electronic health records fit the bill perfectly. They exist in a market context in which the purchaser and user are not the same. EHR buyers are typically health care executives; the users are health care workers. This disconnect results in significantly suppressed innovation-driving feedback mechanisms between the user and vendor relative to the purchaser and vendor.
EHRs are excellent billing and claims management systems, which satisfies their purchasers, but are notorious for their clunky and un-intuitive user interfaces, which negatively affect their users.
One of the original aspirations for electronic health records was that they would launch a digital transformation of U.S. health care by laying the foundation for a deep integration of informatics into all aspects of care delivery. The reality is that they are extremely poor systems for capturing high-quality health and outcomes data and do not have adequate technical infrastructure for supporting downstream data analytics, such as advanced machine learning.
Data from EHRs has generated significant hype as a resource with which to transform evidence-based medicine. But the quality of data from these systems is notoriously poor, and are not immediately suitable for computational analyses, forcing clinical machine learning groups to invest significant time and energy in painstakingly ingesting, cleaning, and vetting data for downstream analysis. And because electronic health record systems are not interoperable — meaning that information in one system is not portable and readily usable by another — the resulting models cannot be deployed and validated across health systems, thereby limiting their potential for clinical impact and raising serious questions about bias and equity.
While the country’s tech giants rely on sophisticated data platforms as foundational elements of their advertising and social media empires, the U.S. health care system has been shackled by software that neither enhances traditional medicine nor paves the way for the future. Without reform, the true potential of artificial intelligence and machine learning in health care will never be realized, nor will the vision for a life sciences medicine continuum in which clinical need and biological discovery are deeply intertwined.
How can ARPA-H change the status quo? Not by funding the development of new EHRs, at least at first. Instead, it should drive the development of tools, operational techniques, and management practices that would allow hospitals to test alternatives on their own systems in real time. In other words, ARPA-H should launch a program to develop platforms for making it easier for a care facility to migrate to a new EHR system. As it stands, once a hospital or health system has committed to one vendor, it is nearly impossible to switch. The transition costs and risks of disrupting patient care are far too high. As a result, large EHR vendors have little reason to innovate.
By substantially decreasing these transition costs, ARPA-H can unleash healthy market forces by encouraging competition between existing vendors and making it viable for newcomers to enter the market. Currently, talented entrepreneurs with ambitions to reform health care are highly reluctant to start EHR companies, despite the need and appetite for alternatives.
What reason is there for believing that such a program could succeed? DARPA program managers are similarly challenged to justify their moonshot projects. With electronic health records there are two clear sets of enabling factors.
The first set includes recent policies, notably the 21st Century Cures Act, to facilitate data exchange between the diverse stakeholders in the U.S. health care system. Patients are already beginning to take advantage of these changes through smartphone apps that allow them to retrieve their health records. Yet there is a vast chasm between an individual patient retrieving a small set of records and an entire hospital or health care system migrating to a new EHR system.
The second set of enabling factors are diverse advances in cloud computing, database technologies, and machine learning (notably massive neural networks in the arenas of transformer-based language models and multi-modal models) that are making it possible to ingest, translate, interpolate, and summarize data from diverse sources. By developing specialized platforms to ingest and standardize all of the data from clinics, hospitals, and health care systems, it should be possible to substantially lower the barrier to switching to a new electronic health record. That way, a new EHR could simply treat the output of an old one as the starting point for a new deployment.
What makes this challenge “ARPA-H Hard,” to borrow a phrase from DARPA parlance, is the precision and efficacy with which it would need to be executed. The technical advances that would underpin this program are already making their way into turn-key solutions — including those specific to health care — by the major cloud vendors and a diverse array of startups. Yet facilitating IT migrations for organizations that are making lifesaving decisions in real time is an entirely different proposition. It would require extraordinary attention to detail, deep understanding of the many legacy systems and medical devices in care delivery settings, and an unparalleled level of operational excellence. It would require building Navy SEAL-caliber teams for health care IT.
These teams are what ARPA-H should aspire to build.
As Collins and his colleagues noted, whereas DARPA is designed to serve a single customer — the Department of Defense — ARPA-H will exist in a complex health care landscape with many stakeholders. Therefore, it should almost be expected that breakthroughs stemming from ARPA-H programs will not only be technological in nature but may also require substantially evolved operational capabilities and management practices.
Time and time again the U.S. has proven itself to be the world leader in the development of advanced science and technology, and yet we regularly settle for less when it comes to the health and well-being of Americans. If it is possible to manage posts, likes, and advertisements for nearly 3 billion users a month; if it is possible to build an everything store and bring one-day delivery to 150 million members; if it is possible to refuel a plane in flight, then it should be possible for a U.S. hospital to run two EHRs simultaneously and choose the best one for its needs.
ARPA-H should lead the way.
Gopal Sarma is a physician-scientist at the Broad Institute of MIT and Harvard, where he leads strategy and operations for its Machine Learning for Health initiative.