The large language models (LLMs) healthcare landscape is rapidly evolving, and both regulators and Health Technology Assessment (HTA) bodies have struggled to maintain pace. In 2019, NICE first published the Evidence Standards Framework (ESF) for digital health technology (DHT) evaluations and updated it in 2022 to cover the types of artificial intelligence (AI) that were most frequently used in the NHS at that time. The ESF provides a set of standards that should be met for the NHS adoption of a DHT, aimed to inform purchasing decisions in the NHS and to guide DHT developers in generating evidence for their technologies. It was designed for the evaluation of DHTs that use AI, are data driven, or have fixed or adaptive machine learning algorithms, including AI image analysis, AI decision support, and health-related chatbots. However, LLMs with healthcare applications are not covered by the ESF, as healthcare LLMs were not available in 2022.
LLMs are generative AI models, trained on extremely large data sets to be effective in various natural language processing tasks. The latest generation of LLMs have sufficient language-handling ability to perform healthcare-related tasks, including text analysis and summarisation, diagnostic assistance, answering medical queries, and image captioning.
LLMs have been identified as having the potential to improve healthcare through improved data handling, process automation, service quality, personalised care, and faster diagnosis. Med-PaLM, Bloom, and LLaMA 3 are current examples of LLMs in healthcare. Med-PaLM provides a range of functions such as diagnostic assistance, synthesising and communicating information from images and other medical data, and discussing results with clinicians through natural language dialogue.
We consider the complexity and requirements for evaluating LLMs and suggest updates to the ESF to help HTA bodies and developers of DHTs meet standards to successfully approach HTA for LLMs in healthcare.
