Mining tumour test results from thousands of research abstracts with AI
IHC-LLMiner uses a language model to find and extract tumour staining profiles from PubMed abstracts, faster and more accurately than GPT-4o on this task.
The question
Pathologists use a lab test called immunohistochemistry, or IHC, to help work out what kind of tumour they are looking at. Which stains show up in which tumours is written across a huge number of research papers, but gathering it by hand is slow.
What we did
We built IHC-LLMiner, a pipeline with two jobs. First it decides which PubMed abstracts are relevant. Then it pulls out the tumour and staining details from the relevant ones. We fine-tuned a language model called Gemma-2 for the task and mapped the results onto standard medical concepts so they can be compared consistently.
What we found
The fine-tuned model got the relevance decision right 91.5% of the time and extracted the profiles correctly 63.3% of the time. It beat GPT-4o on both accuracy and speed.
Why it matters
This points to a way of turning scattered literature into organised, searchable data that researchers and clinicians can use at scale.