Thanks for this nice summary! Also I’m surprised 8% in invasive arm had hematoma/hemorrhage? What

Lots of great questions from the audience!!

McCormick would benefit from airport style moving walkways lol

Being stuck between med/Surg floors and thus forced to adhere to their floor plans when pavilions are built

I’m digging the illustration style

Awesome work!!

This is why I tell the fellows they have to learn r/python/etc so they can work with the data directly!! Being in the weeds will give you perspective you would never get otherwise!

Huge kudos 🙌 to #HBMI #HSIP MD/PhD student @vijeethg.bsky.social , plus Marjorie, Melissa, @keyboardbiologist.bsky.social & the whole #SCRIPT team! 🎉
@iaim-nu.bsky.social
@nufeinbergmed.bsky.social
@nuintmed.bsky.social

3/ Then came the LLMs 🤖✨
GPT-4o, GPT-4o mini, Llama 3.1 🦙, Mixtral-8x7B.
Clear winner: LLMs by far 🚀

To validate externally, we labeled MIMIC notes 📚 and published theses annotations so others can build on this:
🔗 physionet.org/content/immu...

Paper code: github.com/NUSCRIPT/gug...

2/ Most studies lean on structured data (ICD codes 📊) to capture comorbidities.
We tested 10 thresholds → decent performance, but not perfect.

Next: classic NLP 📝
6 bag-of-words classifiers (NB, lasso, ridge, decision tree 🌳, random forest 🌲, SVM).
Results? Also not bad!