My mind is blown by the revelation that prealbumin isn’t a precursor to albumin

As a heterosexual cis white man with an expiring NIH grant, it’s not so great for all of us either

And they're also amazing for rapid iteration of analytical code for computational experiments -- like a 5x speedup at least compared with doing it manually

Have tried using LLM's to help draft reviews and grant applications. Currently they're modestly helpful for turning a brief outline into a first draft, less so for creating that outline or finalizing the draft (still not enough creativity, too much hallucination). But they're improving very fast!

11/ Thanks to the whole clinical trials matching team in KSG at DFCI led by Ethan Cerami, Tali Mazor, James Lindsay, and many others. And thanks to Meta and the Llama Impact Grants program for providing funding to help get this off the ground!

10/ 🚀 We’re still actively refining and improving the pipeline, which will change frequently, but we wanted to share our approach given rapid progress in trial matching and in AI generally. All feedback welcome! #OncologyAI #CancerResearch

9/ Preprint: arxiv.org/abs/2412.17228

8/ Demo app for clinical trial search: huggingface.co/spaces/ksg-d...

7/ 💻 Check it out!
- Code + Models for non-commercial use: huggingface.co/ksg-dfci/Mat...

6/ Results:
-The pipeline achieved precision @ 10 of 89% and mean average precision of 93% for patient-centric queries.
-Embedding models clustered patients by cancer type, aligning clinical trial summaries with patient summaries.
-Real-world application is fast & scalable!

5/ Pipeline overview:
-Condenses and summarizes patient EHR histories using #LLMs
-Extracts trial “spaces” from eligibility criteria on clinicaltrials.gov
-Uses embedding models to rank patient-trial matches
-Double-checks matches with a classifier

4/ Our goal is to help clinicians and ultimately patients identify clinical trials by disease context, not to fully automate screening.

3/ Our Solution: MatchMiner-AI focuses on core “clinical spaces”—combinations of cancer type, treatment history, and biomarkers—for trial matching.

2/ 🧵 The Challenge: Clinical trials are crucial for advancing cancer care, yet many trials fail to recruit enough patients. Existing trial-matching tools often struggle with the complexity of eligibility criteria. #ClinicalTrials