Andrew J. Cowan, MD, is an associate professor in the Clinical Research Division of Fred Hutchinson Cancer Center in Seattle.
One of my favorite pastimes (after hematology, of course) is reading science fiction. My favorite recent series is the Expanse novels, written by James S.A. Corey. For those unfamiliar, the series envisions humanity approximately 300 years in the future, having colonized the inner planets, asteroid belt, and moons of the gas giants. While not explicitly stated, the series hints at advances in medicine that allow artificial intelligence (AI)- and machine learning (ML)-based diagnosis and treatment of many common conditions.
I’m inspired by science fiction, and I believe this series offers a good example of the considerable change we’ll see in the use of AI and ML in medicine in the coming years – perhaps even the next five years. Lately I’ve been thinking about where we are now, how it works, and where this could end up in the next decade.
With respect to AI, you’re likely familiar with large language model (LLM)-based chatbots, the most well-publicized being ChatGPT. ASH Clinical News previously published a rundown of this chatbot,1 so I’m not going to rehash this, but rather point to some ways in which I imagine, and hope, LLMs and AI will broadly impact hematology (and medicine!).
Many people, myself included, sometimes feel like LLMs are a black box, in that you don’t know how the chatbot is generating the information. LLMs use advanced neural networks, which aim to emulate the way the human brain works. These systems rely on machine learning, which is a confusing term that essentially refers to training statistical learning methods on large amounts of so-called “training data.” Those data are used to create a model, which is then processed by another algorithm to create the responses we read in LLMs.
What are some of the ways, both good and bad, that I could see hematology practice being changed by LLMs? Here are a few of my thoughts:
Improved Efficiency
LLMs excel in performing routine tasks that take little or no brain power but consume lots of time. An example is the constant deluge of insurance denials and appeals letters. Many of my colleagues are already using LLMs for assistance with writing appeals, which take time and effort that could be better used for patient care or research.
Another common challenge is documentation in the electronic health record. It’s not farfetched to envision an AI assistant that listens to a clinic visit and accurately transcribes and summarizes the content in the form of a traditional problem-focused note. In fact, efforts to implement this have already begun with Epic and Nuance.2 If this is feasible, it will improve our interactions with patients. Too often, we spend our clinic visits buried in the computer, typing as we talk.
Summarization of the Literature
LLMs also excel at summarizing large volumes of information quickly. Taking advantage of this, an LLM pilot program is being conducted with HemOnc.org,3 where I serve as the editor of the multiple myeloma pages. Thanks to a National Cancer Institute-funded grant, the collaborators are using LLMs to generate human-readable synopses of some of the most relevant anticancer regimens listed on the website, with editorial oversight by the section editors, which is critical. We hope this pilot program will help us better understand how LLMs can be used for the increasingly difficult task of summarizing the complicated development of a cancer treatment regimen, for example.
Clinical Research in Hematology
I was initially drawn to hematology through clinical research. Since embarking on my career, however, I have discovered an unfortunate adverse event: paperwork. So much paperwork. Adverse event grading, writing protocols, responding to institutional review board queries, reviewing serious adverse events, and so on, to name a few. LLMs could potentially streamline clinical research by improving efficiency in many domains.
Take, for example, the process of writing a clinical trial protocol. Summarization of the literature is often a major part of this and can be tedious and time consuming. Using an LLM could increase efficiency in this process, with proper oversight from an experienced clinician investigator. At present, ChatGPT-4 can create a protocol synopsis outline with remarkable accuracy, and even generate a basic statistical design!
Pitfalls
A major concern that has been cited about LLMs and AI is the potential for bias, which should be worrying to hematologists who want to treat patients equally. When we train LLMs, they are exposed to all the collective information on the internet, which means we may inadvertently be perpetuating biases in those data. Concerns have also been raised about LLMs providing nonobjective information about minorities and underrepresented groups. Mitigation strategies for these issues include use of diverse training data and ethical oversight.
One concern I don’t have is that LLMs will replace physician expertise. Although ChatGPT can pass the USMLE step 1 examination with flying colors, we all know that clinical practice is far more than just rote memorization. A personal connection with a patient, I would argue, is a far more important diagnostic tool that helps build trust and enables a thorough diagnostic assessment and discussion of treatment risks and benefits in hematology. I’m confident that AI and LLMs will augment physician practice, not replace physicians.
In summary, AI and LLMs carry great promise for transforming the practice of hematology and medicine in general. It is vital that we engage early on with LLMs to help shape, provide input on, and scrutinize them as they are being developed. I encourage all readers to try the available LLMs and tinker. You’ll be surprised at the potential!
(If you want to read more about this topic, I recommend The AI Revolution in Medicine: GPT-4 and Beyond, written by Peter Lee, Carey Goldberg, and Isaac Kohane, for a deeper dive on how AI could change our practice of medicine.)
Andrew J. Cowan, MD
Associate editor
References
- Lawrence L. “Physician’s AId: will tools like ChatGPT help or harm health care?” ASH Clinical News. 2023;9(9):4-6. https://ashpublications.org/ashclinicalnews/news/7206/Physician-s-AId-Will-Tools-Like-ChatGPT-Help-or.
- Landi H. “Epic, Nuance bring ambient listening, GPT-4 tools to the exam room to help save doctors time.” Fierce Healthcare. July 27, 2023. Accessed September 15, 2023. https://www.fiercehealthcare.com/health-tech/epic-nuance-build-out-more-gpt4-tools-ehrs-help-save-doctors-time.
- Large language model pilot. Updated September 12, 2023. Accessed September 15, 2023. https://www.hemonc.org/wiki/Large_language_model_pilot.org.
The content of the Editor’s Corner is the opinion of the author and does not represent the official position of the American Society of Hematology unless so stated.
Have a comment about this editorial? Let us know what you think; we welcome your feedback. Email the editor your response, along with your full name and professional affiliation if you’d like us to consider publishing it, at ACNEditor@hematology.org.
