You could be forgiven for thinking that A.I. will soon replace human physicians based on headlines such as “The A.I. Doctor Will See You Now,” “Your Future Doctor May Not Be Human,” and “This A.I. Just Beat Human Doctors on a Clinical Exam.” But experts say the reality is more of a collaboration than an ousting: Patients could soon find their lives partly in the hands of A.I. services working alongside human clinicians.
There is no shortage of optimism about A.I. in the medical community. But many also caution the hype surrounding A.I. has yet to be realized in real clinical settings. There are also different visions for how A.I. services could make the biggest impact. And it’s still unclear whether A.I. will improve the lives of patients or just the bottom line for Silicon Valley companies, health care organizations, and insurers.
“I think that all our patients should actually want A.I. technologies to be brought to bear on weaknesses in the health care system, but we need to do it in a non-Silicon Valley hype way,” says Isaac Kohane, a biomedical informatics researcher at Harvard Medical School.
If A.I. works as promised, it could democratize health care by boosting access for underserved communities and lowering costs — a boon in the United States, which ranks poorly on many health measures despite an average annual health care cost of $10,739 per person. A.I. systems could free overworked doctors and reduce the risk of medical errors that may kill tens of thousands, if not hundreds of thousands, of U.S. patients each year. And in many countries with national physician shortages, such as China where overcrowded urban hospitals’ outpatient departments may see up to 10,000 people per day, such technologies don’t need perfect accuracy to prove helpful.
But critics point out that all that promise could vanish if the rush to implement A.I. tramples patient privacy rights, overlooks biases and limitations, or fails to deploy services in a way that improves health outcomes for most people.
“In the same way that technologies can close disparities, they can exacerbate disparities,” says Jayanth Komarneni, founder and chair of the Human Diagnosis Project (Human Dx), a public benefit corporation focused on crowdsourcing medical expertise. “And nothing has that ability to exacerbate disparities like A.I.”
Today, the most popular A.I. techniques are machine learning and its younger cousin, deep learning. Unlike computer programs that rigidly follow rules written by humans, both machine learning and deep learning algorithms can look at a dataset, learn from it, and make new predictions. Deep learning in particular can make impressive predictions by discovering data patterns that people might miss.
But to make the most of these predictions in health care, A.I. can’t go at it alone. Rather, humans still must help make decisions that can have major health and financial consequences. Because A.I. systems lack the general intelligence of humans, they can make baffling predictions that could prove harmful if physicians and hospitals unquestioningly follow them.
The classic example comes from Rich Caruana, a senior researcher at Microsoft Research, as he explained in Engineering and Technology magazine last year. In the 1990s, Caruana worked on a project that tried using an earlier form of machine learning to predict whether a patient with pneumonia was a low-risk or a high-risk case. But trouble arose when the machine learning model tried to predict the case for asthma sufferers, who are high risk because their preexisting breathing difficulties make them vulnerable to pneumonia. The model pegged these patients as low-risk, requiring minor intervention rather than hospitalization — something a human expert never would have done.
If you follow the model blindly, says Kenneth Jung, a research scientist at the Stanford Center for Biomedical Informatics Research, “then you’re hosed. Because the model is saying: ‘Oh, this kid with asthma came in and they got pneumonia but we don’t need to worry about them and we’re sending them home with some antibiotics.’”
Deep-learning predictions can also fail if they encounter unusual data points, such as unique medical cases, for the first time, or when they learn peculiar patterns in specific datasets that do not generalize well to new medical cases.
The A.I. predictions do best when applied to massive datasets, such as in China, which has an advantage in training AI systems thanks to access to large populations and patient data. In February, the journal Nature Medicine published a study from researchers based in San Diego and Guangzhou, China that showed promise in diagnosing many common childhood diseases based on the electronic health records of more than 567,000 children.
But even large datasets can pose problems, particularly when researchers try to apply their algorithm to a new population. In the Nature Medicine study, all the half million patients came from one medical center in Guangzhou, which means there is no guarantee the diagnostic lessons learned from training on that dataset would apply to pediatric cases elsewhere. Each medical center may attract its own unique set of patients — a hospital known for its cardiovascular center, for instance, may attract more critical heart conditions. And findings from a Guangzhou hospital that mostly attracts ethnic Chinese patients may not translate to one in Shanghai with a higher number of foreign-born, non-Chinese patients.
In this 2017 TEDx Talk, Shinjini Kundu of Johns Hopkins Hospital explains how A.I. tools have the potential to glean more from medical images than doctors alone can — including predicting diseases before patients show symptoms.
This extrapolation will prove difficult in other situations as well. For example, says Marzyeh Ghassemi, a computer scientist and biomedical engineer at the University of Toronto, say you have 40,000 ICU patients at Beth Israel Deaconess Medical Center — that’s just one hospital in one city. “And so I have all these papers that have done predictions with this data. Does that work with another hospital in Boston? Maybe. Does it work for a hospital in another state? Would it work in another country? We don’t know.”
While A.I. models may not work in every case, Ghassemi thinks the technology is still worth exploring. “I am very much in favor of taking these models from the bench to the bedside,” she says, “but with really aggressive precautionary steps.”
Those steps need to exist throughout A.I. development and deployment, says I. Glenn Cohen, a law professor at Harvard University and a leader for the Project on Precision Medicine, Artificial Intelligence, and the Law. This may involve verifying the accuracy and transparency of A.I. predictions. And during data collection, researchers will also need to protect patient privacy and ask for consent to use patient data for training A.I.
The consent issue comes up again when the A.I. model is ready for experimental clinical testing with real patients. “Do patients need to be told you’re using the algorithm on them, and does it matter whether the A.I. is completely guiding care or partly guiding care?” Cohen asks. “There is really very little thinking on these questions.”
Ghassemi also advocates for frequently auditing A.I. algorithms to ensure fairness and accuracy across different groups of people based on ethnicity, gender, age, and health insurance. That’s important given how A.I. applications in other fields have already shown that they can easily pick up biases.
After all those steps, the people and companies providing A.I. services will need to sort out legal liability in the case of inevitable mistakes. And unlike most medical devices, which usually need just one regulatory approval, A.I. services may require additional review whenever they learn from new data.
Some regulatory agencies are rethinking how to assess health care A.I. In April, the U.S. Food and Drug Administration (FDA) released a discussion paper to get public feedback about how to update the relevant regulatory review. “What we are continuously trying to do here is get back to our goal of giving people access to technologies, but we’re also realizing that our current methods don’t quite work well,” says Bakul Patel, director for digital health at the FDA. “That’s why we need to look at a holistic approach of the whole product life cycle.”
In addition to issues surrounding access, privacy, and regulations, it also isn’t clear just who stands to benefit the most from A.I. health care services. There are already health care disparities: According to the World Bank and the World Health Organization, half of the globe’s population lacks access to essential health care services and nearly 100 million people are pushed into extreme poverty by health care expenses. Depending on how it is deployed, A.I. could either improve these inequalities, or make them worse.
“A lot of the A.I. discussion has been about how to democratize health care, and I want to see that happening,” says Effy Vayena, a bioethicist at the Federal Institute of Technology in Switzerland.
“If you just end up with a fancier service provision to those who could afford good health care anyway,” she adds, “I’m not sure if that’s the transformation we’re looking for.”
How this all plays out depends on the different visions for implementing A.I. Early development has focused on very narrow diagnostic applications, such as scrutinizing images for hints of skin cancer or nail fungus, or reading chest X-rays. But more recent efforts have tried to diagnose multiple health conditions at once.
In August 2018, Moorfields Eye Hospital in the United Kingdom and DeepMind. the London-based A.I. lab owned by Google’s parent company Alphabet, showed that they had successfully trained an A.I. system to identify more than 50 eye diseases in scans, which matched the performance of leading experts. Similarly broad ambitions drove the San Diego and Guangzhou study that trained A.I. to diagnosis common ailments among children. The latter wasn’t as good at diagnosing pediatric diseases compared to senior physicians, but it did perform better than some junior physicians.
Such A.I. systems may not need to outperform the best human experts to help democratize health care, but simply expand access to current medical standards. Still, so far, many proposed A.I. applications are focused on improving the current standard of care rather than spreading affordable health care around, Cohen says: “Democratizing what we already have would be a much bigger bang for your buck than improving what we have in many areas.”
Accenture, a consulting firm, predicts that top A.I. applications could save the U.S. economy $150 billion per year by 2026. But it’s unclear if patients and health care systems supplemented by taxpayer dollars would benefit, or if more money would simply flow to the tech companies, health care providers, and insurers.
“The question of who is going to drive this and who is going to pay for this is an important question,” says Kohane. “Something a bit hallucinatory about all those business plans is that they think they know how it will work out.”
Even if A.I. services make cost-saving recommendations, human physicians and health care organizations may hesitate to take A.I. advice if they make less money as a result, Kohane cautions. That speaks to the bigger systemic issue of the U.S. health insurers using a fee-for-service model that often rewards physicians and hospitals for adding tests and medical procedures, even when they aren’t needed.
There is another A.I. opportunity that could improve the quality of care while still leaving most medical diagnoses in the hands of doctors. In his 2019 book Deep Medicine, Eric Topol, director and founder of the Scripps Research Translational Institute, talks about creating essentially a supercharged medical Siri — an A.I. assistant to take notes about the interactions between doctors and their patients, enter those notes in electronic health records, and remind physicians to ask about relevant parts of the patient’s history.
“My aspiration is that we decompress the work of doctors and get rid of their data clerk role, help patients take on more responsibility, and key up the data so it doesn’t take so long to review things,” Topol says.
That “never-forgetful medical assistant or scribe,” says Kohane, would require A.I. that can automatically track and transcribe multiple voices between physicians and patients. He supports Topol’s idea, but adds that most A.I. applications in development don’t seem to be focused on such assistants. Still, some companies such as Saykara and DeepScribe have developed services along these lines, and even Google teamed up with Stanford University to test a similar “digital scribe” technology.
An A.I. assistant may sound less exciting than an A.I. doctor, but it could free up physicians to spend more time with their patients and improve overall quality of care. Family physicians in particular often spend more than half of their working days entering data into electronic health records — a main factor behind physical and emotional burnout, which has dire consequences, including patient deaths.
Ironically, electronic health records were supposed to improve medical care and cut costs by making patient information more accessible. Now Topol and many other experts pointed to electronic health records as a cautionary tale for the current hype surrounding A.I. in medicine and health care.
The implementation of electronic health records has already created a patchwork system spread among hundreds of private vendors that mainly succeeds in isolating patient data and makes it inaccessible to both physicians and patients. If history is any guide, many tech companies and health care organizations will feel the pull to follow similar paths by hoarding medical data for their own A.I. systems.
One way around this may be to use a collective intelligence system that aggregates and ranks medical expertise from different sources, says Komarneni, who is trying this approach with Human Dx. Backed by major medical organizations such as the American Medical Association, Human Dx has built an online platform for crowdsourcing advice from thousands of physicians on specific medical cases. Komarneni hopes that such a platform could, in theory, also someday include diagnostic advice from many different A.I. services.
“In the same way that multiple human professionals might look at your case in the future, there is no reason why multiple A.I. couldn’t do it,” Komarneni says.
As doctors wait for their A.I. helpers, crowdsourcing projects like Human Dx “could definitely lead to improved diagnostics or even improved recommendations for therapy,” says Topol, who coauthored a 2018 study on a similar platform called Medscape Consult. The paper concluded collective human intelligence could be a “competitive or complementary strategy” to A.I. in medicine.
But if A.I. services pass all the tests and real-world checks, they could become significant partners for humans in reshaping modern health care.
“There are things that machines will never do well, and then others where they’ll be exceeding what any human can do,” Topol says. “So when you put the two together it’s a very powerful package.”
Jeremy Hsu is a freelance journalist based in New York City. He frequently writes about science and technology for Backchannel, IEEE Spectrum, Popular Science, and Scientific American, among other publications.