RadOnc-GPT, a specialised LLM developed by Mayo Clinic in Arizona is leveraging the open-source Meta Llama 2 model.
By fine-tuning Llama 2 on a large dataset of radiation oncology patient records from Mayo Clinic, RadOnc-GPT has been tailored to understand and process the complex, unstructured data found in this specialised domain. The model was trained locally using a secure GPU server, ensuring that no sensitive patient information was shared outside the institution.
Dr. Wei Liu, Professor of Radiation Oncology and Research Director of Division of Medical Physics at Mayo Clinic in Arizona, emphasises the potential of properly fine-tuned, open-source LLMs in highly specialised healthcare domains like radiation oncology.
The primary goal of RadOnc-GPT is to improve the speed, accuracy, and quality of radiation therapy decision-making, ultimately benefiting both medical practitioners and patients. The model's immediate clinical application lies in patient follow-up, where it can be used to develop a chatbot that answers routine post-radiotherapy questions. This automation reduces the workload of nurses and clinicians, allowing them to focus on more prioritised tasks and direct patient care.
"Efficient tools for language-involved processing can significantly enhance each phase of radiation therapy and potentially improve treatment outcomes," states Dr. Liu, highlighting the time-consuming and variable nature of manually analysing vast amounts of unstructured clinical data.
The Mayo Clinic team plans to expand RadOnc-GPT's capabilities to encompass additional clinical tasks, such as predicting patient outcomes in radiation oncology. They are also considering leveraging the Llama 3 model to further enhance the model's performance.
Curating and preparing the radiation oncology dataset required extensive manual processing, including extracting, separating, and labelling relevant information from patient records. This work was crucial in ensuring the model's accuracy and clinical relevance.