ChatGPT and Large Language Models for Practicing Clinicians: Hype or Hope?

Healthcare is being buffeted by the “4th Industrial Revolution” where artificial intelligence (AI) has become a very real tool for improving diagnosis, predicting clinical trends/outcomes and facilitating redundant tasks that burden clinicians. Yet, the implications for harnessing AI are not straightforward and many balancing measures potentially exist. Asking the important questions related to how we can best leverage this stunning technology without impairing care or accelerating disparities and perpetuating errors is critical. One major element of AI has come to the fore, insinuating itself into mainstream medical conversations. That element is ChatGPT, the most widely used and known of the “Large Language Models” (LLM) released in November 2022 by OpenAI.(1)

The LLMs are massive, powerful machine learning (ML) models that leverage billions of parameters to learn patterns without human guidance. Unlike classic ML models, the LLMs are capable of generating information from inputs, so-called “Generative AI,” and mimic human intelligence in doing so more completely than other model architectures. They are also adept at “transfer learning” where concepts learned in one setting are readily applied to a variety of other tasks. ChatGPT belongs to a subset of LLMs which are called Generative Pre-trained Transformers (GPT) that are trained on massive, complex datasets and information (i.e., text, images, labs, etc.) allowing them to draw on a huge repository of embedded experiences to rapidly answer questions. Recent focus upon ChatGPT, GPT4, Bloom, Stable Diffusion and other LLMs has led to widespread conversations about how they might facilitate clinical care, reduce burn out and churn through common repetitive tasks inherent to medical practice. A major hope is that, through generative AI, clinicians might spend more time with patients and loved ones while spending less time charting in the electronic health record (EHR).

In the recently published book, The AI Revolution in Medicine: GPT-4 and Beyond, Peter Lee, Carey Goldberg and Isaac Kohane take you through their early experiences with beta testing the precursor to GPT4 called “Davinci3.”(2) From the perspective of a computer scientist, journalist and physician-scientist respectively Lee, Goldberg and Kohane explore the functionality of GPT models in healthcare and point towards benefits, potential harms and some level of hype in an honest fashion. They allude to a future in healthcare where LLMs enhance conversations with patients, facilitate clinical encounters and potentially expand access to primary care as chatbots.

But, really, how can LLMs tangibly improve our patient/provider experiences today? Moving beyond the theoretical, several companies are now offering generative AI functionality to improve clinical documentation. One large EHR company is piloting use of ChatGPT in real-world health systems for augmenting physician tasks such as note writing and coding.(3) Similarly, companies have emerged to develop voice assistants which are EHR-vendor agnostic and can securely record and generate notes from “ambient listening” in a clinical encounter, help drive towards optimal billing/coding for the encounter and liaise with the EHR to insert a note for provider review and signature.(4)

What’s next in this space? Given that LLMs are now capable of performing well on medical licensing exams, extracting elements from notes, formatting clinical notes and replying to patient questions among other tasks, we must understand how and where to deploy such tools in our practices.(5) In Creation and Adoption of Large Language Models in Medicine, Shah et al. importantly point out that, as health care providers, we must define optimal uses for these potent tools before technology companies do.(6) Here the authors poignantly remind us that we have had our EHR functionality and healthcare information system structures largely dictated to us and should not allow the same for generative AI. As allergists / immunologists, we do not all need to learn how to code (although it would be a lot cooler if you did!), but we do need to understand how LLMs and future related technologies work, what their limitations are and how they can be optimally trained and used within our specific healthcare environments. In doing so we will better advocate for our patients and our specialty by linking the hard-earned lens of our allergy / immunology domain expertise with bleeding edge technologies.

1. ChatGPT [Internet]. [cited 2023 Aug 22]. Available from: https://openai.com/chatgpt
2. Lee P, Goldberg C, Kohane I. The AI Revolution in Medicine: GPT-4 and Beyond. 1st ed. Pearson Education, Inc.; 2023.
3. Landi H. Fierce Healthcare. 2023 [cited 2023 Aug 22]. Epic is going all in on generative AI in healthcare. Here’s why health systems are eager to test-drive it. Available from: https://www.fiercehealthcare.com/health-tech/epic-moves-forward-bring-generative-ai-healthcare-heres-why-handful-health-systems-are
4. Aquino S. Forbes. [cited 2023 Aug 22]. Health Tech Startup Suki Is Using Artificial Intelligence To Make Patient Records More Accessible To Every Doctor. Available from: https://www.forbes.com/sites/stevenaquino/2023/05/19/health-tech-startup-suki-is-using-artificial-intelligence-to-make-patient-records-more-accessible-to-every-doctor/
5. Lee P, Bubeck S, Petro J. Benefits, Limits, and Risks of GPT-4 as an AI Chatbot for Medicine. N Engl J Med. 2023 Mar 30;388(13):1233–9.
6. Shah NH, Entwistle D, Pfeffer MA. Creation and Adoption of Large Language Models in Medicine. JAMA. 2023 Aug 7


8/29/2023