Enhancing Oncology Care With Federated Learning and Foundation Models

Millions of people worldwide are battling cancer, and personalised care plans are essential for effective diagnosis, treatment, and monitoring of this disease. Recently, Large Language Models (LLMs) have proven valuable in cancer treatment, for instance, extracting key information from Electronic Medical Records (EMRs). This study presents a transformer encoder based LLM, that is domain adapted for Oncology, and outperforms generic models in recognising critical oncology related elements from clinical text. We observe that the development of such domain specific LLMs demands a huge amount of data and computational resources, which is a deterrent to the sustainability development goal of equitable health. To address this problem, we propose a federated learning approach for model development that will eliminate data sharing and centralised computational resource costs. Our evaluations show that the federated approach outperforms the generic base model, highlighting the advantages of collaborative learning in capturing domain specific knowledge and enhancing performance in oncology related NLP tasks. Our work is in line with the United Nations Sustainable Development Goals (SDGs) which are aimed at promoting equitable health and narrowing down the differences in access to advanced cancer treatment.