As artificial intelligence continues to redefine the possibilities within health care, researchers across Dartmouth are creating and applying these new technologies to drive better outcomes for patients and providers alike.
At the second annual Dartmouth on April 18, clinicians, medical researchers, and computer scientists convened to explore the opportunities and complexity that come with modernizing health care in a rapidly evolving digital landscape and to spotlight AI-driven clinical research and innovations by Dartmouth researchers.
“It is really important to recognize that AI tools will augment our existing care models and become a very powerful means for clinicians to advance health care,” Duane Compton, dean of the , said in his welcome address to the audience of over 150 students, researchers, and health providers at the Hanover Inn.
“We have to recognize that there is also a great responsibility to use their power really carefully in the health care setting,” Compton cautioned. “We’re going to see an enormous change in how we conduct health care using these tools and hopefully for the better to improve outcomes for our patients.”
Leading from the forefront by supporting efforts across Dartmouth to build, evaluate, and deploy digital tools that will enhance health care systems is the , launched by Dartmouth in 2023.
“Our last year at CPHAI has been about building a strong foundation, launching collaborative projects, expanding infrastructure, and supporting a diverse community of researchers and students. We are excited to continue shaping the future of ethical, AI-driven health care that is both innovative and inclusive,” said , who presented the center’s annual update and thanked Compton for his invaluable support.
Pranav Rajpurkar, assistant professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School, delivered the Barry D. Pressman, M.D., D’64, M’65, Visiting Lectureship, the event’s first keynote talk.
Rajpurkar, whose research focuses on developing reliable AI systems that can interpret medical data and work alongside clinicians to enhance diagnostic efficiency, presented results from recent studies that advocate a rethink of how AI is combined with physician expertise—not forcing integration, but embracing clear role separation.
Drawing from recent large-scale studies and advances in generalist medical AI systems, he examined promising models where AI and doctors work separately but complementarily, each leveraging their unique strengths.
Navigating the cutting edge of AI and health care
The morning concluded with a panel discussion on the strategies, opportunities, and challenges at Dartmouth and Dartmouth Health for leading AI innovation. Moderated by , chair and professor of biomedical data science, the wide-ranging discussion tackled questions about the integration of AI in the classroom, training faculty, students, and staff at the forefront of AI, implementing AI in patient care, and navigating the cutting edge of AI as an institution.
, the Pat and John Rosenwald Professor in the Department of Computer Science, spoke to the complexity of creating support infrastructure for faculty across the broad range of disciplines at Dartmouth that will likely have different opportunities and challenges in how they adapt to and integrate AI.
Formal training in AI for medical students is “still a work in progress,” said , professor of emergency medicine and chief research officer at Dartmouth Health. Part of it, the panelists agreed, is the ease with which information can now be accessed, prompting educators to rethink how their courses can be designed to teach students how best to utilize new knowledge resources.
One example is the practice of teaching medical students how to write notes after a patient visit, said , chief health information officer at Dartmouth Health. Thanks to emerging AI tools, this mainstay of education may be rendered redundant. “We’re going to have new documentation tools for medical students that can crank out a note for them. It’s a truly brave new world for medical education,” he said.
On the other hand, while applications that use virtual and augmented reality tools in medical training seem cool and show promise, it remains to be seen if they prove truly valuable, said Keith Paulsen, MacLean Professor of Engineering at Thayer School of Engineering.
Exploring ways to integrate AI tools into the health care system while maintaining human oversight and decision-making turned the conversation toward concerns about potential misinformation, liability, and the risk of clinicians relying too heavily on AI, especially when it becomes good enough that people might tend to stop checking and validating outputs carefully.
“My biggest fear around the use of AI in health care would be a mistaken impression on the part of health care leaders that AI can allow you to reduce your workforce. The reality, of course, is AI will never be able to deliver care itself. I think there’s enormous promise in the careful use of AI for diagnostics and prognostics, but the actual on the ground, at the bedside will always be in the province of human beings,” Bernstein said.
For Dartmouth to continue leading from the front, it will be important to identify areas where we have a competitive edge and work across disciplines and create programs that combine the strengths of topic and AI experts, said , professor of epidemiology and senior associate dean of foundational research at Geisel.
Finally, panelists emphasized the importance of protecting patient data while supporting research and innovation, stressing the need for secure computing facilities and proper data governance. “To that point, Dartmouth and Dartmouth Health have partnered on a joint computing facility that is extremely secure, specifically for computing on highly confidential data that have special security constraints,” said Kotz.
AI for clinical impact
The afternoon session was bookended with talks featuring innovations supported by Dartmouth and Dartmouth Health. Katharina Schmolly, a primary care resident physician at Dartmouth Health, showcased zebraMD, a clinical AI assistant that empowers physicians to easily tap into a vast library of rare and genetic disease research to support their clinical practice.
“There is a one in ten risk that you have some kind of rare disease at some point in your lifetime. It’s the same as the risk of getting diabetes, and that’s because there’s over 10,000 rare diseases,” said Schmolly, a veteran U.S. Air Force flight medic who founded zebraMD after witnessing the inequality of care that patients with these conditions can experience.
An early study published by Schmolly and her collaborators showed that their algorithm could have potentially reduced diagnostic delays by two years. The zebraMD app aims to assist physicians in timely diagnosis of rare diseases and provide clinical decision support at the point of care.
Pedram Hosseini, AI lead scientist at LavitaAI, a health care AI company that has collaborated with medical researchers at Dartmouth, spoke about their publicly accessible benchmark for evaluating models that answer medical questions.
High-quality evaluation is a critical component of building any robust AI application, especially in a highly sensitive domain like health care. In his talk, Hosseini reviewed the current state of AI model evaluation in the medical domain such as medical AI assistants and medical question-answering systems.
From consumer medical questions on Lavita’s Medical AI Assist platform, researchers at Lavita AI and Dartmouth have developed a set of criteria that aim to advance long-form medical question-answering using open models and expert-annotated datasets.
Research frontiers
Nigam Shah, chief data scientist for Stanford Health Care, the day’s second keynote speaker, reviewed the use-cases that AI can serve across multiple medical specialties and discussed Stanford Health Care’s efforts to shape the adoption of health AI tools to be useful, reliable, and fair so that they lead to cost-effective and sustainable solutions.
Shah drew on examples from multiple specialties, including pathology, cardiology, surgery, and oncology, to analyze the implications of the choice of business model to ensure the use of AI enhances care quality while managing health care costs.
The AI in Health Care: Current Landscape and Research Frontiers panel moderated by , CPHAI investigator and assistant professor of biomedical data science, focused on research at the intersection of AI and biomedical research currently underway at Dartmouth Health, Geisel, and Dartmouth.
The diverse panel included , associate chief research officer for informatics at Dartmouth Health; David Naeger, chair and professor of radiology at Dartmouth Health; ; , director of genome informatics at Dartmouth Health; and Soroush Vosoughi, assistant professor of computer science.
From using AI to analyze vast quantities of clinical data and developing tools that can reduce physicians’ burnout by assisting with routine clinical tasks to developing novel methods to enhance the transparency of these models and address issues like bias, toxicity, and misalignment through mitigation strategies, their research promises exciting additions to the AI toolbox while keeping cognizant of the limitations and potential issues that need to be addressed while integrating these systems into the real world.
“We heard about a wide spectrum of innovations in AI for precision health,” Hassanpour said in his closing remarks. “These presentations and discussions underscored the incredible potential of artificial intelligence in medicine and demonstrated that our work in this domain is not merely academic, but can be a tool in the hands of physicians, help patients, and serve as a path forward to improve health care systems in our communities.”
