“Clinical reasoning is sort of like a House episode, where we reveal a little bit of information about the patient, they give a differential diagnosis. We reveal a little bit more, they hone their differential diagnosis. Then, they get to a point where they're treating this patient for what they presume to be the diagnosis,” Caruso said.
Since the 1970s, instructors have used actors to teach this skill. The instructor writes a script for the actor to follow, then students interact with the actor and reach a diagnosis. It’s widely considered the best way to teach clinical reasoning, but there are limitations. For one, the actor is typically not a medical professional. If a student’s reasoning process digresses from the scope of the script, the actor will not be able to engage in the same way.
In a more practical sense, using actors is an expensive and time-consuming practice that is difficult to scale up. When institutions don’t have resources to use actors, they can provide written summaries of patient interactions for students to learn from, but these are largely considered inferior to live practice.
In recent years, artificial intelligence simulations have emerged as an alternative to actor simulations, but their versatility with different languages and medical regulations has been limited. Marcos Rojas, who is pursuing his Ph.D. in education at Stanford, is leading the development of Clinical Mind AI, a customizable clinical reasoning tool for use across the globe.
A MEDICAL CHATBOT WITHOUT BORDERS
“For the European community, they decided to do everything in English. That was a problem for us. And the second thing was, all clinical cases were already created, so many clinical cases didn't make sense for our curriculum. A third thing that happened with the platform that is an important detail is that they picked one very specific definition of clinical reasoning, and that's specifically what they were able to measure in this platform and nothing else,” Rojas said. “And that generated many problems in these three things. We don't speak English, those clinical cases are not for our context, and we don't use that definition of clinical reasoning here at this specific university.”
The tool was too rigid, Rojas said, partly owing to limitations in the technology available at the time, and partly owing to the subjectivity of clinical reasoning and the variability of national regulations for specific treatments. For example, he said regulations in Chile on how to treat high blood pressure differed from treatments explained in the tool developed for the European Union.
Rojas said he hopes to make Clinical Mind AI more customizable to account for these differences.
It works like a chatbot, allowing instructors to input patient scenarios and then tailor the simulation to their goals in teaching clinical reasoning, their country’s regulations, et cetera.
It has access to two medical databases: UpToDate and the U.S. Centers for Disease Control and Prevention. Caruso, who is advising Rojas on the project, said an advantage of using AI is that these databases are continuously updated.
“By integrating AI into the clinical reasoning simulations, we will have the power to ensure that the most up-to-date medical knowledge is being taught at any given time, instead of the current state of relying on a physician to constantly remain up to date, which is actually not even possible given the rate of medical discovery right now,” he said.
CUSTOMIZATION, COMPLEXITY AND THE DIGITAL DIVIDE
Rojas said Clinical Mind AI is in the pilot stage, and he's working with educators and students across the globe to make the tool compatible with different languages and instructional goals, in addition to testing the tool’s usability and interface. So far, users have responded well, he said, though some instructors who work with resident-level physicians have requested even more customization.
“Those instructors gave us very important feedback that was, ‘How can I create a clinical case that is more complex?’” Rojas said. “‘Because my residents solve complex clinical cases, and probably also know how to ask questions to the patient, so I don't need the AI simulation. I need the AI to do something else.’”
As a result, he said the beta version gives instructors the option to avoid some types of interactions and skip some activities that were previously included in each simulation. The team is also working on a separate tool to help instructors write scenarios.
Clinical Mind AI is set to launch in 2025. Thinking ahead, when Rojas, Caruso and Shima Salehi, director of the IDEAL research lab at Stanford and another mentor to Rojas, picture a successful rollout, they imagine students learning clinical reasoning in an intuitive way, with results comparable to those from actor simulations. In particular, they hope the tool will help interactive clinical reasoning education scale up and reach institutions with limited resources.
“For Clinical Mind AI, we are looking at how we can use technology to provide to the instructor to enhance their practices, regardless of what kind of resources they have access to and making sure we are serving both underprivileged as well as privileged educational contexts,” Salehi said. “I think it's a unique thing that is not creating a digital divide but augmenting and enhancing practices of educators across different contexts.”
