The American Medical Informatics Association (AMIA) 25×5 Task Force was established in 2022 with the goal of reducing clinician documentation burden by 25 percent in 5 years. Recently, the task force found that “organizational efforts to address documentation burden … were all focused on documentation creation, not on information retrieval.” Similarly, much of the early hype surrounding large language models (LLMs) and generative AI focuses on their ability to produce novel text rather than their potential for information synthesis and retrieval. Such an approach, while useful, misses the mark. Modern physicians do not suffer from a lack of information. Rather, they struggle with a disproportionate signal-to-noise ratio stemming from an overabundance of low-quality information. Instead of using generative AI to produce even more mediocre information, we should apply it to tasks like information retrieval and summarization; doing so will leverage the powers of generative AI to get the right information to the right person at the right time.
The problem
In 2013, Nature reported that biologists had “joined the big-data club” due to advances in technology providing them with massive data sets. In a nation in which the mean patient record is half as long as Shakespeare’s Hamlet, clinicians have also joined this club. Other developing technologies, such as wearables and remote patient monitoring, will likely increase the amount of information physicians are expected to synthesize and act upon.
Much of the funding and conversation surrounding clinician-facing generative AI centers around innovation that, while impactful in the short term, is not substantively innovative. An analysis conducted by GSR Ventures and Maverick Ventures found that while investment in clinician-facing AI has reached a substantial 6.0 billion, “note-taking” was the largest subcategory. While automated note-taking is a worthy and impactful pursuit, this approach glosses over core questions about the purpose of clinical documentation: What should–and should not–be contained in clinical documentation? How should clinical documentation be synthesized and presented to clinicians? Can clinical documentation reasonably satisfy multiple stakeholders with differing priorities (medico-legal, regulatory, billing, quality, etc.)? Automated note-taking solves a short-term problem while ignoring, and perhaps exacerbating, a long-term one. For example, easier document creation could end up resulting in even more stringent and onerous documentation requirements, leading to increased chart bloat and stifling the goal of making less documentation the norm rather than more.
This trend of new technologies digitizing or automating pre-existing processes rather than reimagining them is common–think sending a fax to an email address. Such an approach is practical and can have smooth, fast adoption. However, focusing only on this low-hanging fruit can stifle more substantial innovation and progress. While using generative AI to automate the writing of a prior authorization letter for a physician can save time in the short term, it fails to solve the underlying problem–especially when the insurance company begins to automate the denial letter. More fundamental progress is needed, especially in the field of information synthesis and retrieval.
Revisiting core informatics questions
Imagine that when seeing a new patient, instead of wading through old notes, a physician was presented with a Wikipedia-style document outlining a patient’s summarized medical history, complete with expandable sections and links to more in-depth information. Achieving such a goal requires developing gold standards for what information should be included in such a summary. Thoughtfully revisiting these fundamental questions–What is the right information? Who is the right person? When is the right time?–will strongly influence the ability of generative AI to aid clinicians in substantive ways. Reconsidering these questions could also allow the responsibilities of clinical documentation (medico-legal, regulatory, billing, quality, etc.) to be divided into multiple documents available depending on one’s role. This would allow information that is especially pertinent to clinicians to be emphasized and condensed while information that is more pertinent to other stakeholders would be hidden or deemphasized while still being adequately captured.
More than text
Finally, we wish to point out that medical applications of generative AI and LLMs seem to focus largely on text-based methods of information display. While we are accustomed to information being presented in static blocks of text, this is not necessarily the ideal way to present medical information. There is a plethora of research on what is lost when communication is purely text-based. Embedding generative AI in the electronic health record could further engrain our focus on static-text-based information display. Conversely, generative AI and its real-time capabilities could offer an opportunity to break from this paradigm. Text-based summaries could become dynamic, containing condensed information that can be expanded upon if needed. Information retrieval systems powered by AI could present information in visually intuitive ways, such as a timeline. A patient’s health record could become interactive, allowing question-and-answer queries. We must not let the text-based outputs of generative AI systems limit our conception of these systems’ potential.
Conclusion
We propose that the entry of generative AI into the electronic health record represents an opportunity for informaticists to step back and reconsider our central task of getting the right information to the right person at the right time–including displaying that information in the most intuitive and efficient way. Generative AI must be an opportunity to re-examine the medical record and question assumptions and approaches rather than further ingraining inadequate or outdated approaches. Physicians need tools that help them navigate and gain insight from information, not tools that produce more information to sort through. Let’s ensure that generative AI produces high-quality information that will be helpful to physicians rather than mediocre information that adds to the pile.
Matthew Allen is a medical student.
