ChatCPR provides life-saving guidance during heart-stopping moments

A new study from the University of California San Diego published in the JAMA Internal Medicine Journal showed AI-powered CPR assistance did better than human dispatchers 100% of the time when it came to telling someone how to perform CPR.

The team of scientists from UC San Diego, University of Pittsburgh School of Medicine, Johns Hopkins University and other institutions wanted to see how they could prepare AI models to assist users with CPR, especially with AI’s rise in use.

According to a news release from UC San Diego, Dr. Christopher M. Horvat, director of medical emergency response teams at UPMC Children’s Hospital of Pittsburgh, said, “As new AI technologies emerge, we know people are going to start using them in real-world situations.”

The study aimed to get ahead of those trends and create a reliable AI CPR tool.

First, the team tried to train existing AI models like ChatGPT, Claude and Gemini. The study showed that the bots performed well overall when instructing the basics of CPR, “such as where to press on the chest and how fast to do it.” But it didn’t perform perfectly.

The researchers then developed ChatCPR, an open-source AI model specifically trained to assist callers with CPR. This video demonstration of ChatCPR shows a user calling the AI, explaining the situation and receiving step-by-step directions while they wait for an ambulance to arrive.

ChatCPR was engineered to account for the gap seen in other AI models. The researchers used 911 dispatcher training materials and CPR best practices to teach the AI how to coach during different CPR scenarios.

Then, ChatCPR’s instructions were compared to several real, publicly available 911 calls where CPR instructions were given, and the team found that ChatCPR was more correct than human dispatchers every time. While dispatchers were prone to faltering or misspeaking under stress, ChatCPR gave consistent instruction and was unaffected by the high stakes of the call.

How CPR AI tool fared in comparison

The study reported that dispatchers met 85% of basic CPR guidelines during calls, while ChatCPR met 100%, and dispatchers met 63% of advanced CPR steps while ChatCPR met 99%.

Rema Padman, a study coauthor and professor of Management Science and Healthcare Informatics at Carnegie Mellon University, said: ”No AI system is perfect, but the triangulation of problem identification, deep substantive expertise and AI is the crossroads for healthcare to unlock meaningful breakthroughs."

The news release emphasized that ChatCPR is “a complement, not a replacement, for human responders.” AI is not going anywhere, so researchers aim to improve the powerful tools at society’s fingertips and mitigate potential risks. “The goal is to raise the floor of performance, not to replace trained professionals,” Horvat said.

Perhaps the best solution, however, is to know CPR so you don’t find yourself relying on AI or a dispatcher in a life-or-death situation. According to Nimit Desai, a research affiliate at the UC San Diego Qualcomm Institute, also a medical student at UC San Diego School of Medicine and the lead author of the study, only 2% of Americans are CPR-certified while more than 350,000 Americans experience cardiac arrest outside of a hospital each year. Desai said the survival rate for out-of-hospital cardiac arrest is only 9%.

Real-life application

On May 11, attorney Matt Johnson of Weber County performed CPR on attorney David Winterton after he collapsed mid-court hearing from cardiac arrest, as reported by KSL.com. Johnson had never performed CPR before, but he relied on the CPR skills he learned back in Boy Scouts 45 years prior. He performed CPR until an ambulance arrived 10 minutes later and revived Winterton.

If you don’t want to use models like ChatCPR, it might be time to read up on CPR best practices. Like Johnson, you never know when you might need to save someone’s life.