Lana Garmire, Associate Professor at the University of Michigan, shared a post on LinkedIn:
“Cancer research has never been so personal to me.
Last October, after hearing my father’s passing Anirban Maitra suggested I explore using EHR data for early prediction of pancreatic cancer. Over one year, our team in collaboration with Rui Yin‘s group at the University of Florida, developed a multimodal prediction model called PANCDetect, based on LLMs using three clinical data cohorts, covering nearly 300 million EHR patients.
PANCDetect trained on MarketScan data significantly outperforms other method on prediction accuracy in all time periods, from 6 months to 5 years. The AUROC maintains as high as 0.735, at 5 year prediction. Refined model using UMich EHR data, by adding lab test results, reaches AUROC of over 0.9 in the test set for 5 year prediction.
September 28 2025 marked the first year since my father’s passing from pancreatic cancer. I’m dedicating this work to him.
I hope that one day this model can be deployed in doctors’ offices to make a meaningful impact on patient care. May my father’s spirit rest in peace.
Will share the preprint link as soon as getting approved next week.”
More posts featuring Pancreatic Cancer on OncoDaily.