Top Line: As we’ve been preaching since 2020, anyone hoping to be practicing medicine into the 2030s better be learning to work with–rather than ignoring–rapidly-evolving artificial intelligence.
The Study: Many of those playing around with ChatGPT and the like have learned AI is well-suited to analyze prompts describing detailed patient and tumor characteristics to generate accurate prognostic estimates and treatment recommendations. For instance, asking large language models (LLMs) to estimate the absolute benefit of adding nodal irradiation to breast or prostate volumes for each individual case. Here’s an interesting study that took this process a step further to use GPT-4 from OpenAI to generate a detailed prognostication system for cutaneous squamous cell carcinoma. A major problem with all tumor staging is risk factors are not created equal, especially when continuous variables are arbitrarily dichotomized (e.g. tumor size, depth of invasion, perineural invasion, etc). This study created a prognostic score, called AIRIS (artificial intelligence-derived risk score), through a process called retrieval augmented generation (RAG) enabled generative pretrained transformer (GPT). Yes, the latter part is the GPT in ChatGPT and GPT-4. The key part is RAG. Instead of just asking the LLM (i.e. GPT-4) to create a prognostic system, RAG directs the LLM to first reference and incorporate information from a specified set of documents before generating a response. In this study, the reference documents were 10 large studies describing the impact of cSCC risk factors on outcomes. Through this process, the RAG-enabled GPT-4 model generated AIRIS, which assigned different points to different risk factors to calculate a prognostic class. Not surprisingly, considering what else AI can do, AIRIS proved to discriminate outcomes significantly better than AJCC or Brigham and Women’s Hospital staging systems when validated on a combined prospective and retrospective cohort of 2379 primary cSCC tumors treated between 1996-2023.
TBL: With the right reference sources, RAG-enabled LLMs could prove superior to mere mortals in synthesizing clinical risk factors to estimate oncologic prognosis and risk. | Jairoth, JAMA Dermatol 2025
The Study: Many of those playing around with ChatGPT and the like have learned AI is well-suited to analyze prompts describing detailed patient and tumor characteristics to generate accurate prognostic estimates and treatment recommendations. For instance, asking large language models (LLMs) to estimate the absolute benefit of adding nodal irradiation to breast or prostate volumes for each individual case. Here’s an interesting study that took this process a step further to use GPT-4 from OpenAI to generate a detailed prognostication system for cutaneous squamous cell carcinoma. A major problem with all tumor staging is risk factors are not created equal, especially when continuous variables are arbitrarily dichotomized (e.g. tumor size, depth of invasion, perineural invasion, etc). This study created a prognostic score, called AIRIS (artificial intelligence-derived risk score), through a process called retrieval augmented generation (RAG) enabled generative pretrained transformer (GPT). Yes, the latter part is the GPT in ChatGPT and GPT-4. The key part is RAG. Instead of just asking the LLM (i.e. GPT-4) to create a prognostic system, RAG directs the LLM to first reference and incorporate information from a specified set of documents before generating a response. In this study, the reference documents were 10 large studies describing the impact of cSCC risk factors on outcomes. Through this process, the RAG-enabled GPT-4 model generated AIRIS, which assigned different points to different risk factors to calculate a prognostic class. Not surprisingly, considering what else AI can do, AIRIS proved to discriminate outcomes significantly better than AJCC or Brigham and Women’s Hospital staging systems when validated on a combined prospective and retrospective cohort of 2379 primary cSCC tumors treated between 1996-2023.
TBL: With the right reference sources, RAG-enabled LLMs could prove superior to mere mortals in synthesizing clinical risk factors to estimate oncologic prognosis and risk. | Jairoth, JAMA Dermatol 2025