Better health care through AI? Ambitious but achievable

Isaac S. Kohane, MD, PhD, chair of the Department of Biomedical Informatics at Harvard Medical School, discusses promising uses and the path forward for artificial intelligence (AI) in health care.

Male doctor wearing glasses in front of a bright background with electronic circuit pattern.

In the world of health care, artificial intelligence (AI) is an umbrella term for computer programs that aid medical diagnosis or enhance clinical decision making. So far, AI has proven most successful at analyzing medical images to improve diagnoses in dermatology, ophthalmology and oncology. Some experts estimate that AI could save the U.S. health care system $150 billion annually by 2026. Others caution that real-world challenges such as bias in data collection and ethical concerns remain formidable.

Isaac S. Kohane, MD, PhD, chair of the Department of Biomedical Informatics at Harvard Medical School, is an expert on AI in health care. Dr. Kohane answered questions on the most promising uses of AI, how it could improve doctor-patient interactions and policy challenges that remain daunting.

Interview edited and condensed for clarity

Outside of imaging, what are the most promising patient-use opportunities for AI? How about for the health care delivery or payer space? Are there examples already in use?

AI has proven enormously helpful in interpreting individual genomes. Everyone has millions of genetic variants. AI provides the ability to riffle through vast sources of data to determine the probability that a variant could cause a patient’s disease. For example, Harvard participates in the Undiagnosed Diseases Network (UDN), which uses AI as part of a multidisciplinary evaluation of patients who have seen multiple doctors without getting a diagnosis. The UDN has successfully diagnosed about one-third of participants so far.

However, given that health care spending accounts for more than one-sixth of our GDP, we need to think about the financial impact of AI. A common misperception is that AI in and of itself will lead to cost savings by improving efficiency in diagnosis or care. That won’t happen as long as American health care remains primarily a fee-for-service system, where providers try to maximize reimbursements.

But if system incentives change to payment for outcomes, I have no doubt that AI would save money. These computer programs are fastidious and compulsively aware of the latest trends in patient data, which can reduce expensive, harmful mistakes. Moreover, AI will improve the quality of care, ensuring that it meets the best standard on a consistent basis.

There is great excitement that AI will augment the productivity of doctors, but what would that interface with physician workflow look like? Which potential features will allow physicians to maximize their insights and ability to learn?

We have an impressive example, with electronic medical records (EMRs), of information technology making the patient-doctor experience and the quality of doctors’ lives far worse. The EMR has taken doctors literally from facing their patients to facing computer screens, and from looking for the signs and symptoms of disease and therapeutic response to trying to find the right fields and check boxes to complete on a screen. My colleague Ken Mandl and I warned about this 10 years ago in the New England Journal of Medicine.

In that context, it becomes clear what AI can and should do to improve the quality of medical care. Ideally, an AI program should serve as a faithful, meticulous scribe, listening and watching while a doctor asks questions and does a physical exam. The AI program should then not only create an accurate clinical note describing the clinical encounter but also remind the doctor of any important questions to ask and any diagnostic or therapeutic possibilities to consider.

This would bring doctors back into contact with patients. It would also ensure that an all-too-short clinical encounter is maximally focused on understanding what the patient wants and making sure the appropriate treatment plan is developed.

It will take a while to achieve this, but all the ingredients exist: voice recognition, visual pattern recognition and the ability to leverage data both from the medical literature and large population databases. We need to refine the technology and conduct clinical trials to test it, but the components are there. Bringing them together is ambitious but realizable.

Privacy and re-identification data breaches are increasing in the health care and consumer spaces. Since AI will drive more data collection, do we need a new framework — involving consumers, companies, policy makers — for these issues? How much can we rely on technical data protection solutions versus an ethical and policy framework?

Any technical solution is vulnerable to attack and may fail. There is no question in my mind that we need new policy frameworks that place patient autonomy and self-interest at the center with deliberate consideration of what tradeoffs we are willing to make with regard to patient privacy and disclosure of data.

Much as data giants like Google and Facebook belatedly realized they need to allow customers to decide how much to share and with whom, our health care policy framework must recognize similar levels of autonomy and decision making. One size does not fit all when it comes to patients.

Continue the conversation with us @HMS_ExecEd or with Dr. Kohane @zakkohane.

— Ann MacDonald