When residents at the New York University Grossman School of Medicine (NYU Grossman) interact with a patient, an artificial intelligence (AI) tool can record the conversation and provide feedback on such aspects as the residents’ use of open-ended questions and medical jargon. Residents use that feedback to improve their clinical skills.
At the Johns Hopkins University School of Medicine (Johns Hopkins), students use an AI tool that creates clinical case studies, guides learners through diagnoses and treatment plans, then engages them in a text exchange about their decisions.
Students at the University of Cincinnati (UC) College of Medicine tap into an AI-powered tool for clinical-reason simulation practice and reflective conversations about their clinical encounters with patients, with the tool guiding the students to reach their own conclusions about what they do well and what they should work on more.
At the University of California, San Francisco (UCSF) School of Medicine, students use an AI tool that generates test questions and flash cards based on what’s taught in the students’ classes, rather than on what a public AI tool might find on the internet.
Those are some of the ways that a growing number of medical schools are building AI platforms to carry out precision education. Instruction, feedback, and resources are tailored to the needs, strengths, and learning styles of individual students. While still in early, test-phase iterations — having been deployed within just the past year or two — these tools are providing more customized educational content for students and more individualized feedback than the schools were able to offer before.
While “students love case-based learning, you can only work through so many cases with a live faculty member,” says Maunank Shah, MD, PhD, professor of medicine and epidemiology at Johns Hopkins, in Baltimore.
That points to how AI can address one of the biggest impediments to expanding precision education: staff time.
“It’s really hard to have enough human support to do all the assessment and coaching” that’s required for precision education, notes Jesse Burk-Rafel, MD, MRes, assistant director and vice chair for research at NYU’s Institute for Innovations in Medical Education. Employing AI tools “scales that [capacity] by a factor of 10.”
Another challenge the AI tools seek to address is that, as Shah says, students “go all over the place” on the internet to supplement their classroom instructions with, for instance, case studies and sample test questions and answers. Those public online materials are not vetted by the schools for accuracy, nor are they based on what each school teaches.
“We needed to create our specialized, medically accurate large language model” that could generate materials and feedback that are personalized for each student, says Christy Boscardin, PhD, director of artificial intelligence and student assessment at the UCSF School of Medicine. The school built a tool that aligns with UCSF curriculum and each student’s learning goals.
Below are examples of how students are using new personalized AI tools at four medical schools. Although the tools all perform several tasks, these examples focus on a few selected functions. Important questions remain about quality, accuracy, and long-term impact.
Patient interactions
Communication Compass, an AI-empowered toolkit created by NYU Grossman, analyzes residents’ verbal interactions with patients. Those interactions, picked up by the tool through ambient audio, are still observed by coaches and clinicians. But the analysis of how the residents talk — producing data about such things as word choice and sentence construction — “is hard for humans to do, because it’s so hard to measure,” Burk-Rafel notes.
The feedback shows residents areas for improvement, he says. “Maybe they’re using complex language that might not be great for patients with low health literacy.”
Residents can give the toolkit their notes from the patient visits to get feedback on those as well. That function was requested by residents, he adds, because “we didn’t have enough teachers and other staff to read those notes.”
At UCSF a tool called Curate provides coaching to third-year students during their clinical visits with patients. As described by UCSF, the tool aids in the development of clinical reasoning by providing feedback on each student’s illness scripts and also their notes, using medically accurate content.
Customized study materials
Johns Hopkins and UCSF employ slightly different approaches to using AI to create study and practice materials for students. One critical factor is that the tools have been trained to primarily use materials from each school and its own students.
First-year medical students at Johns Hopkins can use an AI-powered system called PRISM (Precision Review and Interactive Simulation in Medicine), which taps into content and related materials from a course: Organ Systems Foundations of Medicine. It uses that content to create patient case studies, summary notes of lectures on specific topics, and sample questions, answers, and flash cards, according to the introduction provided to students.
Students can ask for case studies on specific syndromes or organisms, or request a random case. The tool walks students through clinical-reasoning steps, interpretation of diagnostic tests, and reviews of core pharmacology concepts.
Shah notes that PRISM is designed “to act like a JHU clinical preceptor” in tone, reasoning, and pedagogy.
At UCSF, the Curate tool organizes material from the students’ own classroom notes as well as texts of lectures, then creates personalized flash cards, illness scripts, and differential diagnoses to help the students identify both strengths and areas to work on.
The system links students to PDFs of materials from the school’s curriculum and curated outside material, providing vetted content that allows students to dive deeper into specific topics, Boscardin says. The tool is used by students in the school’s Bridges Curriculum, which emphasizes integrated learning on foundational sciences, clinical skills, inquiry, and systems improvements.
Working though issues with an AI coach
At the UC College of Medicine, third- and fourth-year students can tap into an interactive AI coach called CAR-E (Coaching with AI-Reinforced Education) to talk about matters such as their clinical encounters and residency plans. Students type or dictate information, thoughts, and questions into the web-based system, which remembers that information so it can continually learn about the student and create customized, fine-tuned responses.
CAR-E gives guidance and asks probing questions, using coaching protocols that it has been trained on, in order to help students reflect, identify gaps in their knowledge, and make decisions, says Laurah Turner, PhD, MS, associate dean for artificial intelligence and educational informatics.
Students have used it to review their clinical-care experiences, as well as to help apply for residencies, she says.
The system initially frustrates some students, Turner adds, “because it won’t just give them an answer.” Eventually, she says, students say that working through the questions gave them a better understanding of how to proceed.
Turner recalls a student who was preparing for residency interviews and wanted to know what to ask during interviews to help her decide which institutions would offer the best fit for her. A public tool like ChatGPT would suggest questions, she notes. CAR-E did not.
“It asked her questions like, ‘What does she value in a team?’ ‘What does she value in locations and geography?’ ‘In the experience of her residency?’ She had to work through all of that.
“She found that at the end it was valuable, because then she actually did a deep dive into what she wants in a residency program.”
Moving with caution
As with just about all use of AI, issues arise about the privacy of the information it receives and the accuracy of the content it provides.
The schools try to mitigate the risks of exposing student and patient data to people unauthorized to see it, and of feeding students inaccurate medical information, by building customized in-house systems. (Those systems are often built on existing AI tools, such as ChatGPT.) The systems are typically closed rather than public, with access limited to registered users. The schools establish rules about what individual information can and cannot be uploaded into the tools, in order to adhere to the Health Insurance Portability and Accountability Act (HIPPA), which protects patient information held by health entities, and the Family Educational Rights and Privacy Act (FERPA), which protects student information retained by schools.
For accuracy, the systems are trained on and link to content that is based at or vetted by the schools, although some of the systems also scour certain outside sources. To varying degrees, faculty at the schools assess the accuracy of the content used or produced by the AI.
Moving forward, the schools plan to continually revise and expand the tools based on learner use and feedback. That will require more systematic, data-driven assessments. For now, faculty members look at some student usage data and get responses from individuals, which appears to be overwhelmingly positive.
Written student comments at UCSF include, “I like that I can trust that Curate uses more trustworthy sources that are vetted by UCSF” and “I like how it breaks down complex topics into easy to understand ways. I wish professors used it to teach, because sometimes they don’t really explain difficult subjects well.”
As they expand these tools, faculty say they must critically examine where AI brings significant added value to medical education, and where it might just be something of a toy.
One risk is that AI could dampen the quality of a student’s learning. NYU’s Burk-Rafel refers to the phenomenon of “never skilling,” where someone does not learn a skill because they learned to rely on technology to do the work for them.
“There are ways to do harm if a learner is given access to a tool that takes away a productive struggle, a struggle that led to learning, a struggle that led to growth that was formative,” Burk-Rafel says. “If you shortcut that, you do not learn in the same way.”
Those are among the reasons that schools are “trying to be thoughtful and judicious about how this is done,” says Shah at Johns Hopkins. “We don’t want to incorporate AI for the sake of incorporating AI.”
