How to prove AI’s clinical value before payment models reflect it
Key takeaways
Administrative AI is spreading because it fits the documentation and billing demands of fee-for-service, while clinical AI still faces adoption barriers, according to a new Peterson Health Technology Institute report. PHTI recommends focusing clinical AI on conditions such as hypertension and mental health, where access is limited and outcomes are measurable, and it offers a four-step framework for proving value before outcomes-based payment arrives. A separate American Medical Association framework for agentic AI holds that clinical authority should be earned, with oversight and accountability adjusted as evidence grows.
As healthcare continues to deploy artificial intelligence, two recent frameworks stress the importance of ensuring AI implementation doesn’t replicate age-old problems of misaligned incentives and inefficient workflows.
Administrative uses for AI are catching on largely because they align with the emphasis on documentation and billing so common in fee-for-service models, the Peterson Health Technology Institute (PHTI) noted in its latest report. Clinical applications, meanwhile, “continue to face significant barriers to adoption, despite growing evidence of their value.”
To start, the research organization said the industry should focus clinical AI on health conditions where access is limited and outcomes are measurable. Here, guidance from the American Medical Association (AMA) enters the picture to spell out why “clinical authority should be earned” for agentic AI tools.
Opportunities for mental health, hypertension care
Though administrative AI may create value for individual institutions, PHTI pointed out that it’s poised to increase overall spending as billable activity increases. (It doesn’t help that provider and payer AI-assisted coding tools increasingly disagree about the complexity of care delivered.)
On the other hand, mounting clinical evidence suggests that AI can help manage hypertension by risk-stratifying patients, managing medication, recommending treatment decisions and more. Hypertension is a good candidate for AI-supported care because of its “well-established and readily measurable endpoints,” namely lower blood pressure, and as a result “creates a pathway toward more proactive and scalable population management.”
Mental health presents another opportunity, PHTI said. Outcomes are admittedly more complex, but existing scales for measuring mental health symptoms and functions also offer clinical endpoints. What’s more, purpose-built and validated chatbots can support engagement, monitoring, intervention and escalation (as necessary) at a scale mental health professionals cannot replicate.
Proving AI’s clinical value before payment reform happens
Outcomes-based payment models represent the starting point for the industry’s shift toward clinical AI adoption. “Organizations are more likely to invest in this work,” PHTI said, “when there is a clear path to compensation for outcomes and cost savings generated through AI-based care.”
That is some time off. Plus, outcomes-based models aren’t one-size-fits-all, and it will take effort to determine what works best for a given condition or population. In the meantime, PHTI provided a four-step framework for proving clinical AI’s value.
- Understand what works, as assessing outcomes and performance helps users and purchasers gain confidence in AI tools.
- Set rules for coverage and use that define scope of practice, clinical liability and insurance coverage.
- Purchase effectively by evaluating AI vendors against established outcomes. This will set the stage for designing outcomes-based payment models.
- Redesign care delivery to embed clinical AI where appropriate and reframe roles and responsibilities for clinical staff.
Deciding what AI is permitted to do, and when
The AMA framework notes that AI’s evolution means tools can perform “increasingly consequential clinical functions,” and it is meant to help organizations understand what systems should do for whom.
The process begins with defining the clinical function and evaluating model performance, safety, clinical benefits and targeted patient population against the potential for the tool to make errors. That will help organizations determine what clinical AI systems should be permitted to do under what conditions, as well as who should be responsible for the risk of using such a system.
The AMA stressed that this is a continuous process. Authority, oversight and accountability are all bound to change as evidence shows that AI can do more. No matter how care delivery and payment models advance, these factors “must remain aligned in service of better patient care.”
Brian Eastwood is a Boston-based writer with more than 10 years of experience covering healthcare IT and healthcare delivery. He also writes about enterprise IT, consumer technology, and corporate leadership.