Category: Artificial Intelligence

AI is getting pretty darn good at patient diagnosis challenges… but don’t bother asking it to show its work.

A new study in npj Digital Medicine pitted GPT-4V against human physicians on 207 image challenges designed to test the reader’s ability to diagnose a patient based on a series of pictures and some basic clinical background info.

• Researchers at the NIH and Weill Cornell Medicine then asked GPT-4V to provide step-by-step reasoning for how it chose the answer.

• Nine physicians then tackled the same questions in both a closed-book (no outside help) and open-book format (could use outside materials and online resources).

How’d they stack up?

• GPT-4V and the physicians both scored high marks for accurate diagnoses (81.6% vs. 77.8%), with a statistically insignificant difference in performance. 

• GPT-4V bested the physicians on the closed-book test, selecting more correct diagnoses.

• Physicians bounced back to beat GPT-4V on the open-book test, particularly on the most difficult questions.

• GPT-4V also performed well in cases where physicians answered incorrectly, maintaining over 78% accuracy.

Good job AI, but there’s a catch. The rationales that GPT-4V provided were riddled with mistakes – even if the final answer was correct – with error rates as high as 27% for image comprehension.

The Takeaway

There could easily come a day when clinical AI surpasses human physicians on the diagnosis front, but that day isn’t here quite yet. Real care delivery also doesn’t bless physicians with a set of multiple choice options, and hallucinating the rationale behind diagnoses doesn’t cut it with actual patients.

When it rains it pours for AI research, and a trio of studies published just last week suggest that many new generative AI tools might not be ready for prime time with patients.

The research that grabbed the most headlines came out of UCSD, finding that GenAI-drafted replies to patient messages led to more compassionate responses, but didn’t cut down on overall messaging time.

• Although GenAI reduced the time physicians spent writing replies by 6%, that was more than offset by a 22% increase in read time, while also increasing average reply lengths by 18%.

• Some of the physicians were also put off by the “overly nice” tone of the GenAI message drafts, and recommended that future research look into “how much empathy is too much empathy” from the patient perspective.

Another study in Lancet Digital Health showed that GPT-4 can effectively generate replies to health questions from cancer patients… as well as replies that might kill them.

• Mass General Brigham researchers had six radiation oncologists review GPT-4’s responses to simulated questions from cancer patients for 100 scenarios, finding that 58% of its replies were acceptable to send to patients without any editing, 7% could lead to severe harm, and one was potentially lethal.

• The verdict? Generative AI has the potential to reduce workloads, but it’s still essential to “keep doctors in the loop.”

A team at Mount Sinai took a different path to a similar conclusion, finding that four popular GenAI models have a long way to go until they’re better than humans at matching medical issues to the correct diagnostic codes.

• After having GPT-3.5, GPT-4, Gemini Pro, and Llama2-70b analyze and code 27,000 unique diagnoses, GPT-4 came out on top in terms of exact matches, achieving an uninspiring accuracy of 49.8%.

The Takeaway

While it isn’t exactly earth-shattering news that GenAI still has room to improve, the underlying theme with each of these studies is more that its impact is far from black and white. GenAI is rarely completely right or completely wrong, and although there’s no doubt we’ll get to the point where it’s working its magic without as many tradeoffs, this research confirms that we’re definitely not there yet.

Medical AI is on the brink of improving outcomes for countless patients, prompting a trio of all-star researchers to pen an NEJM AI article tackling what might be its biggest obstacle: real-world adoption.

What drives real-world adoption? Those who have been around the block as many times as Dr. Michael Abramoff, Dr. Tinglong Dai, and Dr. James Zou are all-too familiar with the answer… Reimbursement makes the world go ‘round.

To help medical AI developers get their tools in front of the patients who need them, the authors explore the pros and cons of current paths to reimbursement, while offering novel frameworks that could lead to better financial sustainability.

Traditional Fee-for-Service treats medical AI similarly to how new drugs or medical devices are reimbursed, and is a viable path for AI that can clear the hurdle of demonstrating improvements to clinical outcomes, health equity, clinician productivity, and cost-effectiveness (e.g. AI for diabetic eye exams).

• Meeting these criteria is a prerequisite for adopting AI in healthcare, yet even among the 692 FDA-authorized AI systems, few have been able to pass the test. The approach carries substantial risk in terms of time and resources for AI developers.

• Despite those limitations, FFS might be appropriate for AI because health systems are adept at assessing the financial impact of new technologies under it, and reimbursement through a CPT code provides hard-to-match financial sustainability.

Value-based care frameworks provide reimbursement on the basis of patient- or population-related metrics (MIPS, HEDIS, full capitation), and obtaining authorization for medical AI to “count” toward closing care gaps for MIPS and HEDIS has been shown to be considerably more straightforward than attaining a CPT code.

• That said, if a given measure is not met (e.g. 80% of the population must receive an annual diabetic eye examination), the financial benefit of closing even three quarters of that care gap is typically zero, potentially disincentivizing AI adoption.

Given the limitations of existing pathways, the authors offer a potential new approach that’s derived from the Medicare Part B model, which reimburses drugs administered in an outpatient setting based on a “cost plus” markup.

• Here, providers could acquire the rights to use AI, then get reimbursed based on the average cost of the service plus a specified margin, contingent upon CMS coverage of a particular CPT code.
• This model essentially splits revenue between AI creators and users, and would alleviate some of the tensions of both FFS and VBC models.

The Takeaway

Without sustainable reimbursement, widespread medical AI adoption won’t be possible. Although the quest continues for a silver bullet (even the authors’ revenue-sharing model still carries the risk of overutilization and requires the creation of new CPT codes), exploring novel approaches is essential given the challenges of achieving reimbursement through existing FFS and VBC pathways.

The New England Journal of Medicine’s just-released NEJM AI publication is off to the races, with its February issue including a stellar breakdown of how academic medical centers are managing the influx of predictive models and AI tools.

Researchers identified three governance phenotypes for managing the AI deluge:

• Well-Defined Governance – health systems have explicit, comprehensive procedures for the evaluation of AI and predictive models.
• Emerging Governance – systems are in the process of adapting previously established approaches for things like EHRs to govern AI.
• Interpersonal Governance – a small team or single person is tasked with making decisions about model implementation without consistent evaluation requirements. 

Regardless of the phenotype, interviews with AI leadership at 13 academic medical centers revealed that chaotic implementations are hard to avoid, partly due to external factors like vague regulatory standards.

• Most AI decision makers were aware of how the FDA regulates software, but believed those rules were “broad and loose,” and many thought they only applied to EHRs and third party vendors rather than health systems.

AI governance teams report better adherence to new solutions that prioritize limiting clicks for providers when they’re implemented. Effective governance of prediction models requires a broader approach, yet streamlining workflows is still a primary consideration for most implementations. That’s leading to trouble down the road considering predictive models’ impact on patient care, health equity, and quality care.

The Takeaway

Even well-equipped academic medical centers are struggling to effectively identify and mitigate the countless potential pitfalls that come along with predictive AI implementation. Existing AI governance structures within healthcare orgs all seem to be in need of additional guidance, and more guardrails from both the industry and regulators might help turn AI ambitions into AI-improved outcomes.