For the Symposium on The Law And Policy Of AI, Robotics, and Telemedicine In Health Care.
How will artificial intelligence (AI) change medicine? AI, powered by “big data” in health, promises to transform medical practice, but specifics remain inchoate. Reports that AI performs certain tasks at the level of specialists stoke worries that AI will “replace” physicians. These worries are probably overblown; AI is unlikely to replace many physicians in the foreseeable future. A more productive set of questions considers how AI and physicians should interact, including how AI can improve the care physicians deliver, how AI can best enable physicians to focus on the patient relationship, and how physicians should review the recommendations and predictions of AI. Answering those questions requires clarity about the larger function of AI: not just what tasks AI can do or how it can do them, but what role it will play in the context of physicians, other patients, and providers within the overall medical system. Medical AI can improve for patients and improve the practice of medicine for providers—as long as its development is supported by an understanding of what role it can and should play. Four different roles each have the possibility to be transformative for providers and patients: AI can push the frontiers of medicine; it can replicate and democratize medical expertise; it can automate medical drudgery; and it can allocate medical resources.
Pushing frontiers
The headline promise of artificial intelligence is that it will push the frontiers of medical care, enabling a level of precision and personalization previously unavailable. Some medical AI aims to use underlying patterns of patient characteristics—biological, social, or environmental—to tailor treatment. In this role, AI increases provider capabilities and can result in better patient care. For instance, AI can power “artificial pancreas” devices that combine a continuous glucose monitor, a subcutaneous insulin pump, and a controller. The AI enables the controller to “learn” the glucose response of a particular patient and to tailor insulin dosage to keep glucose levels within safe limits over the course of the day. Ideally, AI-controlled artificial pancreases can perform better than earlier technology because the patterns of individual response are complex and differ from patient to patient. Similarly, AI can improve the cutting-edge practice of medicine in various situations where identifying illness, making prognoses, or suggesting treatment depends on complex networks of factors not easily captured in explicit practice guidelines or other sources of medical knowledge. Medical AI could also push frontiers of explicit knowledge by suggesting new hypotheses and avenues of research, whether about medical care pathways or about the development of new pharmaceuticals.
Replicating and democratizing expertise
Despite the glamour of pushing medical frontiers, much work in medical AI
has focused on replicating existing human capacity, such as AI that identifies
malignant skin lesions from skin images as well as dermatologists. The IDx-DR device exemplifies this
pattern. Approved by the FDA in May
2018 for autonomous diagnosis of diabetic retinopathy, IDx-DR consists of
an automated retinal fundus imaging camera combined with two AI algorithms. A minimally trained operator uses the camera
to acquire retinal fundus images for a diabetic patient; these images are sent
to a first algorithm that evaluates image quality; if the quality is adequate,
the images are evaluated by a second algorithm to determine whether the patient
has more-than-mild diabetic retinopathy, indicating a need for further care. IDx-DR can allow a primary-care physician—or
other provider—to
provide screening care that would otherwise require an ophthalmologist.
This role for AI is most connected to worries about physician
replacement. But we are far from that
possibility. A much closer possibility
is the use of medical AI to replicate capabilities that are now limited to
relatively small numbers of specialists—and thus, to make those capabilities
available to a much broader set of patients.
Rather than AI replacing physicians who are present, AI could help
non-specialists provide patients with care they might not otherwise be able to
access.
Hundreds of millions of patients across the world lack meaningful access
to even primary care; billions lack access to specialists of various kinds. AI will not replace ophthalmologists, but it
could allow patients without ready access to ophthalmologists to get screenings
they would otherwise not receive. Google
is pursuing
the possibility of deploying such ophthalmological AI in drastically
underserved populations in rural India. Similarly, AI-powered screening of
potentially malignant skin lesions via smartphone
cameras will not replace existing dermatologists, but could help patients who
lack ready access know that they need to seek out care. If AI can be deployed to increase the
capacity of whatever providers are present—or to allow patients access to the
technology on their own—it could radically improve the care available,
especially to underserved patient populations.
Automating drudgery
AI can also transform medical practice in a more quotidian way, by automating
medical drudgery. Some of this ability
has been recognized; for instance, AI can triage
radiologic images and flag those that demand closer attention, freeing
physician time.
But AI could also make a difference in a broad set of time-consuming
tasks: paperwork. Too much of
physicians’ time is occupied by routine data search, data entry, and other desk
work. According to one
study, physicians spend nearly half of their time overall on desk work and
electronic health record work; even in the examination room, about a third of
physician time is spent on similar tasks.
AI has the potential to automate at least some of this work, using
natural language processing to identify the most relevant information from
patient history or useful articles and resources from the medical
literature. Even nearer-term, AI-driven
speech recognition and natural language processing could take notes on patient
encounters, structure the data, and flag the most relevant points. These interventions, while not exciting on
their face, could free substantial physician time to interact more with
patients, improving the practice of medicine and decreasing the leading cause
of physician burnout. They could also
enhance accuracy and patient privacy by automating the otherwise error-prone
process of entering structured data or transcribing notes.
Allocating resources
Finally, AI offers the possibility of helping allocate resources
efficiently across the medical system.
Rather than freeing up provider time, or extending or replicating
provider capabilities, AI can direct existing scarce resources where they are
likely to do the most good. For
instance, a resource-triage AI could help allocate scarce inpatient hospital
beds when patient demand exceeds availability; similarly, AI could help
allocate drugs during shortages (both issues with ethical and
legal implications). More simply, AI
could help optimize staffing to better match patterns of patient need. While these uses of AI raise ethical
questions—who deserves resources, and what is to prevent administrators from
using AI to maximize revenue rather than effectiveness of care?—resource
allocation AI is nevertheless likely to become a pervasive, behind-the-scenes
aspect of the medical system.
Conclusion
Medical AI will not replace physicians, at least not in the foreseeable future. But it will dramatically shift the shape of medical practice. When developers, providers, patients, policymakers, and regulators are debating how AI should be developed, regulated, and brought into practice, it is important to be clear about not only what medical AI can do and how it does it, but exactly what role it will play in the medical system to come.
W. Nicholson Price II is Assistant Professor at the University of Michigan Law School. You can reach him by e-mail at wnp at umich.edu