The HIMSS Physician Committee understands the ever-evolving and the complex circumstances physicians face when it comes to Generative AI, specifically Large Language Models (LLMs), in their practices. To try to help bring some of this information into perspective, they have teamed up with industry subject matter experts to create a whitepaper, which will be published in six separate sections to make the content more digestible.
The six sections will cover: 1) High-level overview of LLMs; 2) Early-Phase Risks of Implementing Language Models in Healthcare; 3) How LLMs can aid in documentation; 4) Securities for data and privacy; 5) Issues and concerns with LLMs; and 6) What organizations should be doing, now, to prepare and train their staff.
This is section one, and the succeeding sections will be published over the next several weeks. We hope you find these educational and beneficial.
As a side note, the Physician Committee also held a panel discussion titled Unlocking Large Language Models in Healthcare.
A large language model is a type of artificial intelligence (AI) model that is designed to process and generate human-like language. These models are built using deep learning techniques, particularly with architectures like transformer models, and they leverage vast amounts of textual data to learn patterns and relationships within language.
The main purpose of large language models is natural language processing (NLP). They can understand and generate human language, perform various NLP tasks such as text classification, sentiment analysis, language translation, question-answering, summarization, and more. These models have a wide range of applications across different industries, including customer support, content generation, language translation, chatbots, and even assisting with coding and software development.
The term "large" in "large language model" refers to the vast number of parameters or weights in the model. The number of parameters determines the model's capacity and capability to understand complex language structures and generate coherent and contextually relevant responses. The more parameters a model has, the larger and more powerful it generally is. Models are typically trained on public data sourced from the internet. Although this includes a wide range and depth of information, it can also include biased or factually inaccurate data or may have key gaps as noted below.
Some well-known examples of large language models include OpenAI's GPT series (e.g., ChatGPT), Google's BARD, and Facebook’s LLaMa. While these models perform well on many use cases reflected in their training data, they currently have substantial limitations in the context of healthcare. A key challenge is a lack of publicly available electronic medical record data suitable for model training and tuning, given that the Health Insurance Portability and Accountability Act of 1996 (HIPAA) limits the sharing or selling of patient healthcare information (PHI). Additionally, there is a lack of established benchmarks and testing frameworks for evaluating the performance of Clinical Language Models (CLaMs, source). Anticipated future advances in model training methodologies, such as privacy preserving training and deidentification, will be needed to develop and test high-quality CLaMs suitable for the most complex applications.
We, the authors, believe that while LLMs have great potential in healthcare, equally great care must be taken when identifying appropriate use cases, evaluating performance, and safeguarding sensitive patient health information. In this document, we will discuss the following: various uses for LLMs in healthcare and their relative risk; specific discussions around LLMs and clinical documentation; security and privacy considerations; organizational planning for AI and LLMs; and issues of concern with LLMs in healthcare. The overarching message is that while LLMs are a powerful tool with the potential to improve the patient and provider experience, care and consideration are essential to leveraging them in a safe, appropriate, and equitable manner.
The views and opinions expressed in this content or by commenters are those of the author and do not necessarily reflect the official policy or position of HIMSS or its affiliates.