Session 4
Chair: Edward H. (Ted) Shortliffe, MD, PhD, MACP, FACMI, FAIMBE, FIAHSI (Columbia University)
Speaker: Robert A. (Bob) Greenes, MD, PhD, (Arizona State University)
Professor Emeritus, College of Health Solutions
Title: AI and the Spectrum of Clinical Decision Support
Abstract: This talk will describe an expanded view of what we mean by CDS in the era of AI. We have already broadened this definition considerably over the years from the ideas of prediction models, diagnostic assistance, therapy guidance, and guidelines/prescriptive rules, to include order sets and document forms as ways of facilitating data reporting and report review, and the use of visualization to make relationships more apparent. Now with the AI era, more opportunities exist for capturing and analyzing content and interacting with the provider and patient to facilitate optimal decision making and actions.
Speaker: Randolph A. (Randy) Miller, MD (Vanderbilt University)
Professor Emeritus, Department of Biomedical Informatics
Title: Truth Underlying Clinical Diagnosis: Can LLMs ever know it?
Abstract: Until three decades ago, patient-diagnostician encounters were recorded via clinicians’ often highly descriptive handwritten notes. Subsequently, electronic health record systems have degraded the information content of such notes. Similarly, accurate information on how disorders affect patients comes from careful review of the peer-reviewed literature, not from LLM’s statistical summaries of millions of low fidelity EMR case records. The patient charts used by LLMs lack detailed and consistent disease and finding names, and only record random subsets of all patient findings. When new disorders arise, clinicians can recognize them before they become noticeable in LLMs. In summary, LLMs face more handicaps in knowledge representation and understanding than careful literature reviews by expert clinicians.
Speaker: Lucila Ohno-Machado, MD, PhD (Yale University)
Chair, Department of Biomedical Informatics and Data Science
Title: Risk Assessment and Prognosis: Can AI Do It Better Than We Do (and How Will We Know)?
Abstract: Estimating the probability of disease development and progression constitutes one of the main uses of AI in medicine, but how do we know that a model is producing the right estimates? Furthermore, how do we know that these estimates are applicable to populations that are different from the ones used to train the model so we can feel confidence using the model? Generative AI has re-opened the discussion on the need to consider the quality of AI model estimates in the context of individuals, even though evaluation measures are still primarily focused on the population as a whole.
Speaker: Mark Musen, MD, PhD (Stanford University)
Professor, Biomedical Informatics Research
Director, Stanford Center for Biomedical Informatics Research
Title: Dumbing Down or Getting Real? The Changing Agenda for Clinical Decision-Support Systems
Abstract: Developers of decision-support systems in the 1970s had the goal of developing reasoning systems that could behave as master clinicians did. There was an emphasis on making inferences about clinical data over time, on using anatomic knowledge to aid diagnosis, on understanding how diseases might spread, and on exploiting explicit representations of evidence-based guidelines to reason about optimal care. Current decision-support systems tend not to do any of these things. What have we lost and what have we gained with the shift in our agenda and in the changes in our underling technology? Are we taking advantage of approaches that are more scalable than anything that we have we have ever seen before, or are we looking for our keys where the light is?
