More than two-thirds of U.S. physicians have changed their mind about generative AI and now view it as beneficial to healthcare. But as AI grows more powerful and prevalent in medicine, apprehensions remain high among medical professionals.
For the last 18 months, I’ve examined the potential uses and misuses of generative AI in medicine; research that culminated in the new book ChatGPT, MD: How AI-Empowered Patients & Doctors Can Take Back Control of American Medicine. Over that time, I’ve seen the concerns of clinicians evolve—from worries about AI’s reliability and, consequently, patient safety to a new set of fears: Who will be held liable when something goes wrong?
From safety to suits: A new AI fear emerges
Technology experts have grown increasingly certain that next-gen AI technologies will prove vastly safer and more reliable for patients, especially under expert human oversight. As evidence, recall that Google’s first medical AI model, Med-PaLM, achieved a mere “passing score” (>60%) on the U.S. medical licensing exam in late 2022. Five months later, its successor, Med-PaLM 2, scored at an “expert” doctor level (85%).
Since then, numerous studies have shown that generative AI increasingly outperforms medical professionals in various tasks. These include diagnosis, treatment decisions, data analysis and even expressing empathy.
Despite these technological advancements, errors in medicine can and will occur, regardless of whether the expertise comes from human clinicians or advanced AI.
Fault lines: Navigating AI’s legal terrain
Legal experts anticipate that as AI tools become more integrated into healthcare, determining liability will come down to whether errors result from AI decisions, human oversight or a combination of both.
For instance, if doctors use a generative AI tool in their offices for diagnosing or treating a patient and something goes wrong, the physician would likely be held liable, especially if it’s deemed that clinical judgement should have overridden the AI’s recommendations.
But the scenarios get more complex when generative AI is used without direct physician oversight. As an example, who is liable when patients rely on generative AI’s medical advice without ever consulting a doctor? Or what if a clinician encourages a patient to use an at-home AI tool for help interpreting wearable device data, and the AI’s advice leads to a serious health issue?
In a working paper, legal scholars from the universities of Michigan, Penn State and Harvard explored these challenges, noting: “Demonstrating the cause of an injury is already often hard in the medical context, where outcomes are frequently probabilistic rather than deterministic. Adding in AI models that are often nonintuitive and sometimes inscrutable will likely make causation even more challenging to demonstrate.”
That paper, published earlier this year in the New England Journal of Medicine, is based on hundreds of software-related tort cases and offers insights into the murky waters of AI liability, including how the courts might handle AI-related malpractice cases.
However, Mello pointed out that direct case law on any type of AI model remains “very sparse.” And when it comes to liability implications of using generative AI, specifically, there’s no public record of such cases being litigated.
“At the end of the day, it has almost always been the case that the physician is on the hook when things go wrong in patient care,” she noted but also added, “As long as physicians are using this to inform a decision with other information and not acting like a robot, deciding purely based on the output, I suspect they’ll have a fairly strong defense against most of the claims that might relate to their use of GPTs.”
She emphasized that while AI tools can improve patient care by enhancing diagnostics and treatment options, providers must be vigilant about the liability these tools could introduce. To minimize risk, she recommends four steps.
Understand the limits of AI tools: AI should not be seen as a replacement for human judgment. Instead, it should be used as a supportive tool to enhance clinical decisions.
Negotiate terms of use: Mello urges healthcare professionals to negotiate terms of service with AI developers like Nvidia, OpenAI, Google and others. This includes pushing back on today’s “incredibly broad” and “irresponsible” disclaimers that deny any liability for medical harm.
Apply risk assessment tools: Mello’s team developed a framework that helps providers assess the liability risks associated with AI. It considers factors like the likelihood of errors, the potential severity of harm caused and whether human oversight can effectively mitigate these risks.
Stay informed and prepared: “Over time, as AI use penetrates more deeply into clinical practice, customs will start to change,” Mello noted. Clinicians need to stay informed as the legal landscape shifts.
The high cost of hesitation: AI and patient safety
While concerns about the use of generative AI in healthcare are understandable, it’s critical to weigh these fears against the existing flaws in medical practice.
Each year, misdiagnoses lead to 371,000 American deaths while another 424,000 patients suffer permanent disabilities. Meanwhile, more than 250,000 deaths occur due to avoidable medical errors in the United States. Half a million people die annually from poorly managed chronic diseases, leading to preventable heart attacks, strokes, cancers, kidney failures and amputations.
Our nation’s healthcare professionals don’t have the time in their daily practice to address the totality of patient needs. That’s because the demand for medical services is higher than ever at a time when health insurers—with their restrictive policies and bureaucratic requirements—make it harder than ever to provide excellent care. Generative AI can help.
But it is imperative for policymakers, legal experts and healthcare professionals to collaborate on a framework that promotes the safe and effective use of this technology. As part of their work, they’ll need to address concerns over liability. Ultimately, they must recognize that the risks of not using generative AI to improve care will far outweigh the dangers posed by the technology itself. Only then can our nation reduce the enormous human toll resulting from our current medical failures.
Pay attention to the media coverage around artificial intelligence, and it’s easy to get the sense that technologies such as chatbots pose an “existential crisis” to everything from the economy to democracy.
These threats are real, and proactive regulation is crucial. But it’s also important to highlight AI’s many positive applications, especially in health care.
Consider the Mayo Clinic, the largest integrated, nonprofit medical practice in the world, which has created more than 160 AI algorithms in cardiology, neurology, radiology and other specialties. Forty of those have already been deployed in patient care.
To better understand how AI is used in medicine, I spoke with John Halamka, a physician trained in medical informatics who is president of Mayo Clinic Platform. As he explained to me, “AI is just the simulation of human intelligence via machines.”
Halamka distinguished between predictive and generative AI. The former involves mathematical models that use patterns from the past to predict the future; the latter uses text or images to generate a sort of human-like interaction.
It’s that first type that’s most valuable to medicine today. As Halamka described, predictive AI can look at the experiences of millions of patients and their illnesses to help answer a simple question: “What can we do to ensure that you have the best journey possible with the fewest potholes along the way?”
For instance, let’s say someone is diagnosed with Type 2 diabetes. Instead of giving generic recommendations for anyone with the condition, an algorithm can predict the best care plan for that patient using their age, geography, racial and ethnic background, existing medical conditions and nutritional habits.
This kind of patient-centered treatment isn’t new; physicians have long been individualizing recommendations. So in this sense, predictive AI is just one more tool to aid in clinical decision-making.
The quality of the algorithm depends on the quantity and diversity of data. I was astounded to learn that the Mayo Clinic team has signed data-partnering agreements with clinical systems across the United States and globally, including in Canada, Brazil and Israel. By the end of 2023, Halamka expects the network of organizations to encompass more than 100 million patients whose medical records, with identifying information removed, will be used to improve care for others.
Predictive AI can also augment diagnoses. For example, to detect colon cancer, standard practice is for gastroenterologists to perform a colonoscopy and manually identify and remove precancerous polyps. But some studies estimate that 1 in 4 cancerous lesions are missed during screening colonoscopies.
Predictive AI can dramatically improve detection. The software has been “trained” to identify polyps by looking at many pictures of them, and when it detects one during the colonoscopy, it alerts the physician to take a closer look. One randomized controlled trial at eight centers in the United States, Britain and Italy found that using such AI reduced the miss rate of potentially cancerous lesions by more than half, from 32.4 percent to 15.5 percent.
Halamka made a provocative statement that within the next five years, it could be considered malpractice not to use AI in colorectal cancer screening.
But he was also careful to point out that “it’s not AI replacing a doctor, but AI augmenting a doctor to provide additional insight.” There is so much unmet need that technology won’t reduce the need for health-care providers; instead, he argued, “we’ll be able to see more patients and across more geographies.”
Generative AI, on the other hand, is a “completely different kind of animal,” Halamka said. Some tools, such as ChatGPT, are trained on un-curated materials found on the internet. Because the inputs themselves contain inaccurate information, the models can produce inappropriate and misleading text. Moreover, whereas the quality of predictive AI can be measured, generative AI models produce different answers to the same question each time, making validation more challenging.
At the moment, there are too many concerns over quality and accuracy for generative AI to direct clinical care. Still, it holds tremendous potential as a method to reduce administrative burden. Some clinics are already using apps that automatically transcribe a patient’s visit. Instead of creating the medical record from scratch, physicians would edit the transcript, saving them valuable time.
Though Halamka is clearly a proponent of AI’s use in medicine, he urges federal oversight. Just as the Food and Drug Administration vets new medications, there should be a process to independently validate algorithms and share results publicly. Moreover, Halamka is championing efforts to prevent the perpetuation of existing biases in health care in AI applications.
This is a cautious and thoughtful approach. Just like any tool, AI must be studied rigorously and deployed carefully, while heeding the warning to “first, do no harm.”
Nevertheless, AI holds incredible promise to make health care safer, more accessible and more equitable.
Of 26 health systems surveyed by MedCity News, nearly half used automated tools to respond to the Covid-19 pandemic, but none of them were regulated. Even as some hospitals continued using these algorithms, experts cautioned against their use in high-stakes decisions.
A year ago, Michigan Medicine faced a dire situation. In March of 2020, the health system predicted it would have three times as many patients as its 1,000-bed capacity — and that was the best-case scenario. Hospital leadership prepared for this grim prediction by opening a field hospital in a nearby indoor track facility, where patients could go if they were stable, but still needed hospital care. But they faced another predicament: How would they decide who to send there?
Two weeks before the field hospital was set to open, Michigan Medicine decided to use a risk model developed by Epic Systems to flag patients at risk of deterioration. Patients were given a score of 0 to 100, intended to help care teams determine if they might need an ICU bed in the near future. Although the model wasn’t developed specifically for Covid-19 patients, it was the best option available at the time, said Dr. Karandeep Singh, an assistant professor of learning health sciences at the University of Michigan and chair of Michigan Medicine’s clinical intelligence committee. But there was no peer-reviewed research to show how well it actually worked.
Researchers tested it on over 300 Covid-19 patients between March and May. They were looking for scores that would indicate when patients would need to go to the ICU, and if there was a point where patients almost certainly wouldn’t need intensive care.
“We did find a threshold where if you remained below that threshold, 90% of patients wouldn’t need to go to the ICU,” Singh said. “Is that enough to make a decision on? We didn’t think so.”
But if the number of patients were to far exceed the health system’s capacity, it would be helpful to have some way to assist with those decisions.
“It was something that we definitely thought about implementing if that day were to come,” he said in a February interview.
Thankfully, that day never came.
The survey Michigan Medicine is one of 80 hospitals contacted by MedCity News between January and April in a survey of decision-support systems implemented during the pandemic. Of the 26 respondents, 12 used machine learning tools or automated decision systems as part of their pandemic response. Larger hospitals and academic medical centers used them more frequently.
Faced with scarcities in testing, masks, hospital beds and vaccines, several of the hospitals turned to models as they prepared for difficult decisions. The deterioration index created by Epic was one of the most widely implemented — more than 100 hospitals are currently using it — but in many cases, hospitals also formulated their own algorithms.
They built models to predict which patients were most likely to test positive when shortages of swabs and reagents backlogged tests early in the pandemic. Others developed risk-scoring tools to help determine who should be contacted first for monoclonal antibody treatment, or which Covid patients should be enrolled in at-home monitoring programs.
MedCity News also interviewed hospitals on their processes for evaluating software tools to ensure they are accurate and unbiased. Currently, the FDA does not require some clinical decision-support systems to be cleared as medical devices, leaving the developers of these tools and the hospitals that implement them responsible for vetting them.
Among the hospitals that published efficacy data, some of the models were only evaluated through retrospective studies. This can pose a challenge in figuring out how clinicians actually use them in practice, and how well they work in real time. And while some of the hospitals tested whether the models were accurate across different groups of patients — such as people of a certain race, gender or location — this practice wasn’t universal.
As more companies spin up these models, researchers cautioned that they need to be designed and implemented carefully, to ensure they don’t yield biased results.
An ongoing review of more than 200 Covid-19 risk-prediction models found that the majority had a high risk of bias, meaning the data they were trained on might not represent the real world.
“It’s that very careful and non-trivial process of defining exactly what we want the algorithm to be doing,” said Ziad Obermeyer, an associate professor of health policy and management at UC Berkeley who studies machine learning in healthcare. “I think an optimistic view is that the pandemic functions as a wakeup call for us to be a lot more careful in all of the ways we’ve talked about with how we build algorithms, how we evaluate them, and what we want them to do.”
Algorithms can’t be a proxy for tough decisions Concerns about bias are not new to healthcare. In a paper published two years ago, Obermeyer found a tool used by several hospitals to prioritize high-risk patients for additional care resources was biased against Black patients. By equating patients’ health needs with the cost of care, the developers built an algorithm that yielded discriminatory results.
More recently, a rule-based system developed by Stanford Medicine to determine who would get the Covid-19 vaccine first ended up prioritizing administrators and doctors who were seeing patients remotely, leaving out most of its 1,300 residents who had been working on the front lines. After an uproar, the university attributed the errors to a “complex algorithm,” though there was no machine learning involved.
Both examples highlight the importance of thinking through what exactly a model is designed to do — and not using them as a proxy to avoid the hard questions.
“The Stanford thing was another example of, we wanted the algorithm to do A, but we told it to do B. I think many health systems are doing something similar,” Obermeyer said. “You want to give the vaccine first to people who need it the most — how do we measure that?”
The urgency that the pandemic created was a complicating factor. With little information and few proven systems to work with in the beginning, health systems began throwing ideas at the wall to see what works. One expert questioned whether people might be abdicating some responsibility to these tools.
“Hard decisions are being made at hospitals all the time, especially in this space, but I’m worried about algorithms being the idea of where the responsibility gets shifted,” said Varoon Mathur, a technology fellow at NYU’s AI Now Institute, in a Zoom interview. “Tough decisions are going to be made, I don’t think there are any doubts about that. But what are those tough decisions? We don’t actually name what constraints we’re hitting up against.”
The wild, wild west There currently is no gold standard for how hospitals should implement machine learning tools, and little regulatory oversight for models designed to support physicians’ decisions, resulting in an environment that Mathur described as the “wild, wild west.”
How these systems were used varied significantly from hospital to hospital.
Early in the pandemic, Cleveland Clinic used a model to predict which patients were most likely to test positive for the virus as tests were limited. Researchers developed it using health record data from more than 11,000 patients in Ohio and Florida, including 818 who tested positive for Covid-19. Later, they created a similar risk calculator to determine which patients were most likely to be hospitalized for Covid-19, which was used to prioritize which patients would be contacted daily as part of an at-home monitoring program.
Initially, anyone who tested positive for Covid-19 could enroll in this program, but as cases began to tick up, “you could see how quickly the nurses and care managers who were running this program were overwhelmed,” said Dr. Lara Jehi, Chief Research Information Officer at Cleveland Clinic. “When you had thousands of patients who tested positive, how could you contact all of them?”
While the tool included dozens of factors, such as a patient’s age, sex, BMI, zip code, and whether they smoked or got their flu shot, it’s also worth noting that demographic information significantly changed the results. For example, a patient’s race “far outweighs” any medical comorbidity when used by the tool to estimate hospitalization risk, according to a paper published in Plos One. Cleveland Clinic recently made the model available to other health systems.
Others, like Stanford Health Care and 731-bed Santa Clara County Medical Center, started using Epic’s clinical deterioration index before developing their own Covid-specific risk models. At one point, Stanford developed its own risk-scoring tool, which was built using past data from other patients who had similar respiratory diseases, such as the flu, pneumonia, or acute respiratory distress syndrome. It was designed to predict which patients would need ventilation within two days, and someone’s risk of dying from the disease at the time of admission.
Stanford tested the model to see how it worked on retrospective data from 159 patients that were hospitalized with Covid-19, and cross-validated it with Salt Lake City-based Intermountain Healthcare, a process that took several months. Although this gave some additional assurance — Salt Lake City and Palo Alto have very different populations, smoking rates and demographics — it still wasn’t representative of some patient groups across the U.S.
“Ideally, what we would want to do is run the model specifically on different populations, like on African Americans or Hispanics and see how it performs to ensure it’s performing the same for different groups,” Tina Hernandez-Boussard, an associate professor of medicine, biomedical data science and surgery at Stanford, said in a February interview. “That’s something we’re actively seeking. Our numbers are still a little low to do that right now.”
Stanford planned to implement the model earlier this year, but ultimately tabled it as Covid-19 cases fell.
‘The target is moving so rapidly’ Although large medical centers were more likely to have implemented automated systems, there were a few notable holdouts. For example, UC San Francisco Health, Duke Health and Dignity Health all said they opted not to use risk-prediction models or other machine learning tools in their pandemic responses.
“It’s pretty wild out there and I’ll be honest with you — the dynamics are changing so rapidly,” said Dr. Erich Huang, chief officer for data quality at Duke Health and director of Duke Forge. “You might have a model that makes sense for the conditions of last month but do they make sense for the conditions of next month?”
That’s especially true as new variants spread across the U.S., and more adults are vaccinated, changing the nature and pace of the disease. But other, less obvious factors might also affect the data. For instance, Huang pointed to big differences in social mobility across the state of North Carolina, and whether people complied with local restrictions. Differing social and demographic factors across communities, such as where people work and whether they have health insurance, can also affect how a model performs.
“There are so many different axes of variability, I’d feel hard pressed to be comfortable using machine learning or AI at this point in time,” he said. “We need to be careful and understand the stakes of what we’re doing, especially in healthcare.”
Leadership at one of the largest public hospitals in the U.S., 600-bed LAC+USC Medical Center in Los Angeles, also steered away from using predictive models, even as it faced an alarming surge in cases over the winter months.
At most, the hospital used alerts to remind physicians to wear protective equipment when a patient has tested positive for Covid-19.
“My impression is that the industry is not anywhere near ready to deploy fully automated stuff just because of the risks involved,” said Dr. Phillip Gruber, LAC+USC’s chief medical information officer. “Our institution and a lot of institutions in our region are still focused on core competencies. We have to be good stewards of taxpayer dollars.”
When the data itself is biased Developers have to contend with the fact that any model developed in healthcare will be biased, because the data itself is biased; how people access and interact with health systems in the U.S. is fundamentally unequal.
How that information is recorded in electronic health record systems (EHR) can also be a source of bias, NYU’s Mathur said. People don’t always self-report their race or ethnicity in a way that fits neatly within the parameters of an EHR. Not everyone trusts health systems, and many people struggle to even access care in the first place.
“Demographic variables are not going to be sharply nuanced. Even if they are… in my opinion, they’re not clean enough or good enough to be nuanced into a model,” Mathur said.
The information hospitals have had to work with during the pandemic is particularly messy. Differences in testing access and missing demographic data also affect how resources are distributed and other responses to the pandemic.
“It’s very striking because everything we know about the pandemic is viewed through the lens of number of cases or number of deaths,” UC Berkeley’s Obermeyer said. “But all of that depends on access to testing.”
At the hospital level, internal data wouldn’t be enough to truly follow whether an algorithm to predict adverse events from Covid-19 was actually working. Developers would have to look at social security data on mortality, or whether the patient went to another hospital, to track down what happened.
“What about the people a physician sends home — if they die and don’t come back?” he said.
Researchers at Mount Sinai Health System tested a machine learning tool to predict critical events in Covid-19 patients — such as dialysis, intubation or ICU admission — to ensure it worked across different patient demographics. But they still ran into their own limitations, even though the New York-based hospital system serves a diverse group of patients.
They tested how the model performed across Mount Sinai’s different hospitals. In some cases, when the model wasn’t very robust, it yielded different results, said Benjamin Glicksberg, an assistant professor of genetics and genomic sciences at Mount Sinai and a member of its Hasso Plattner Institute for Digital Health.
They also tested how it worked in different subgroups of patients to ensure it didn’t perform disproportionately better for patients from one demographic.
“If there’s a bias in the data going in, there’s almost certainly going to be a bias in the data coming out of it,” he said in a Zoom interview. “Unfortunately, I think it’s going to be a matter of having more information that can approximate these external factors that may drive these discrepancies. A lot of that is social determinants of health, which are not captured well in the EHR. That’s going to be critical for how we assess model fairness.”
Even after checking for whether a model yields fair and accurate results, the work isn’t done yet. Hospitals must continue to validate continuously to ensure they’re still working as intended — especially in a situation as fast-moving as a pandemic.
A bigger role for regulators All of this is stirring up a broader discussion about how much of a role regulators should have in how decision-support systems are implemented.
Of the hospitals surveyed by MedCity News, none of the models they developed had been cleared by the FDA, and most of the external tools they implemented also hadn’t gone through any regulatory review.
“My experience suggests that most models are put into practice with very little evidence of their effects on outcomes because they are presumed to work, or at least to be more efficient than other decision-making processes,” Kellie Owens, a researcher for Data & Society, a nonprofit that studies the social implications of technology, wrote in an email. “I think we still need to develop better ways to conduct algorithmic risk assessments in medicine. I’d like to see the FDA take a much larger role in regulating AI and machine learning models before their implementation.”
Developers should also ask themselves if the communities they’re serving have a say in how the system is built, or whether it is needed in the first place. The majority of hospitals surveyed did not share with patients if a model was used in their care or involve patients in the development process.
In some cases, the best option might be the simplest one: don’t build.
In the meantime, hospitals are left to sift through existing published data, preprints and vendor promises to decide on the best option. To date, Michigan Medicine’s paper is still the only one that has been published on Epic’s Deterioration Index.
Care teams there used Epic’s score as a support tool for its rapid response teams to check in on patients. But the health system was also looking at other options.
“The short game was that we had to go with the score we had,” Singh said. “The longer game was, Epic’s deterioration index is proprietary. That raises questions about what is in it.”
HENRY KOTULA 