Researchers at Johns Hopkins APL are working on a proof of concept for a conversational artificial intelligence agent that will be able to provide medical guidance to untrained soldiers in plain English, by applying knowledge gleaned from established care procedures. (U.S. Air National Guard photo by Staff Sgt. Sara Kolinski)

Soldiers on the battlefield can find themselves suddenly thrust into unfamiliar territory, caring for wounded comrades whose lives may depend on their ability to render quick and effective aid.

Researchers at the Johns Hopkins Applied Physics Laboratory in Laurel are working on a proof of concept for a conversational artificial intelligence agent that can give the guidance soldiers need in situations like this.

For anyone who remembers the television series “Chuck,” it can be likened to a relatively simplified version of the Intersect which provides the protagonist with the knowledge necessary to overcome dangerous situations and defeat international terrorists and other stock villains.

Known as Clinical Practice Guideline-driven AI, the APL project is based on a type of AI known as a large language model. It’s similar to ChatGPT, though not associated or affiliated with that model.

Complicated system

Sam Barham, a computer scientist in APL’s Research and Exploratory Development Department, is leading the CPG-AI project.

Breaking down the work behind the project, he explained that upward of 20 or 30 individual components running behind the scenes are needed to enable a conversational agent to render assistance.

This includes “everything from search components to deciding which information from the search is relevant to managing the structure of the dialogue,” Barham said.

“Our system is able to answer medical questions,” he explained. “That sounds like a simple task, but it has to get broken down into as many tiny atomic subtasks as you can.”

Some of these subtasks get handed to a large language model, while others are handed to a traditional software component.

“The LLM can search through a bunch of clinical practice guidelines that the Army has written,” Barham said. “If the user doesn’t know how to do something, the model can disambiguate the information it provides. The LLM and traditional software systems work together in a complex network, with information flowing between them.”

Timely research

Amanda Galante, who oversees the Assured Care research portfolio at APL, said connecting emerging technology with an urgent military need is both timely and more important than ever.

Barham and his team “are applying models like ChatGPT to solve sponsor problems, which presents considerable challenges,” she said. “How can we harness these powerful tools, while also ensuring accuracy, as well as transparency? We need a capability that can provide relevant knowledge in a usable manner.”

To date, the researchers have converted more than 30 clinical practice guidelines from the Department of Defense Joint Trauma System to be ingested as text by their model, including guidelines for treating burns, blunt trauma, and other common conditions encountered by warfighters on the battlefield.

In the project’s first phase, Barham and his team produced a prototype model that can infer a patient’s condition based on conversational input, answer questions accurately and without jargon, and guide the user through the care algorithms for tactical field care addressing the most common battlefield injuries, including breathing issues, burns and bleeding.

Closer to reality

As other researchers continue to refine artificial intelligence and the ability for conversational AI agents to respond appropriately, APL researchers believe the reality of the concept is within reach.

“Three years ago this was the realm of science fiction,” Barham said. “Two years ago, some researchers began to sense that this was going to be possible. A year ago it wasn’t really clear, but the Lab believes in it.”

“The big question for us is whether it’s even feasible,” Gallante said. “Does it make sense to try to provide medical care in this context using LLMs? It’s still a research question at this point.”

The biggest challenge ahead, Barham said, is making the architecture smaller.

“We’re using laptops now,” he noted. “What the ultimate form factor looks like requires input from the sponsor. it could be a cellphone, or whatever hardware a soldier may be wearing in battle. The challenge now is to shrink it down so it can run on something smaller. To me, it seems really feasible now.”

If it does work, Galante said, it could potentially be adapted for other uses.

“Sam’s work is part of our vision for the future where we are trying to achieve this big goal of having health care anytime, anywhere,” she said. “The intent would be to develop technology that would be sensible to the sorts of applications we see that could be useful to anyone who is in a resource-limited environment.”