John Ayers believes most healthcare AI hype has not yet translated into meaningful patient impact — though he thinks that may soon change. This belief drove Ayers and a team of researchers to create ChatCPR, an AI agent launched this week that coaches users through CPR in real time.
Ayers, head of AI at the University of California San Diego’s Altman Clinical and Translational Research Institute, is the lead author of a widely discussed 2023 JAMA study that found AI chatbots’ responses to patient messages are often more accurate and empathic than those written by human doctors. His prominence in healthcare AI research has made him a frequent commentator on television news segments — and Ayers said one on-air moment from earlier this year really stuck with him, forcing him to grapple with AI’s true effect on patient outcomes.
“The reporter asked me, ‘Isn’t this exciting, that AI is going to save lives?’ And I’m like, ‘What are you talking about?’ It’s not really helping anybody yet — it’s all hype, no reality. The way it’s being delivered is all on the back end,” he said.
So, Ayers and his research team at UC San Diego teamed up with other researchers from organizations including Johns Hopkins and UPMC to see if AI can start to truly save lives. They looked at where seconds matter most and existing systems fail: out-of-hospital cardiac arrest.
Each year, more than 350,000 Americans suffer cardiac arrests outside of hospitals, and nearly 90% do not survive. However, only 2% of Americans are CPR-certified, Ayers pointed out. When someone collapses, most people just call 911 and wait for emergency responders to arrive.
“The reality is that waiting costs you. Every minute it takes to deliver CPR, the efficacy of it is reduced,” Ayers declared.
The team of researchers benchmarked major AI models — including OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini — on CPR instruction tasks. Most of the models did reasonably well on basic CPR guidance, but often missed more nuanced but clinically important instructions.
So, the researchers built ChatCPR to handle more advanced, guideline-critical details. A study published Monday in JAMA not only introduced the tool, but also showed that it outperformed 911 dispatchers in guiding bystanders through CPR when tested against recordings from real 911 calls.
The research team rolled out ChatCPR this week as an open-source public resource rather than a commercial product. They are making the training materials, guidelines, prompts and architecture publicly available so that the right companies and emergency-response organizations can build on it, improve it and deploy it broadly, Ayers said.
In his eyes, the key challenge in healthcare AI is implementation — not necessarily having the most advanced model. This is why the team intentionally built ChatCPR on a relatively small, lower-performing language model and still achieved strong results through careful design and domain-specific training.
Ayers said this means the tool could eventually run directly on smartphones without requiring internet connectivity.
He also noted that ChatCPR could potentially reduce disparities in emergency care access.
“CPR should not be a luxury good. But even in the wealthiest country in the world, depending on where you’re at and the resources around you for emerging medical services, it is often a luxury,” Ayers remarked.
He hopes open-sourcing the tool will accelerate its adoption and improvement across the healthcare system.
Photo: dzm1try, Getty Images
