Narges Razavian

Assistant Prof @NYULangone. #AIInMedicine.
Interested in AI, fairness, medicine and science. Prev @LTIatCMU, @CILVRatNYU.

All views my own and #woman_life_freedom

Narges Razaviannarges@sigmoid.social
2023-05-06

RT @togethercompute
The first RedPajama models are here! The 3B and 7B models are now available under Apache 2.0 license, including instruction-tuned and chat versions!
This project demonstrates the power of the open-source AI community with many contributors ... 🧵 together.xyz/blog/redpajama-mo

Narges Razaviannarges@sigmoid.social
2023-05-06

RT @vitaliychiley
Our team at @mosaicml has been working on releasing something special:

We're proud to announce that we are OPEN SOURCING a 7B LLM trained to 1T tokens
The MPT model outperforms ALL other open source models!
Code: github.com/mosaicml/llm-foundr
Blog: mosaicml.com/blog/mpt-7b

🧵

Narges Razaviannarges@sigmoid.social
2023-05-06

RT @simocristea
New preprint on scGPT: first foundation large language model for single cell biology, pretrained on 10 million cells.

Just as text is made of words, cells are characterized by genes

Some thoughts on how cool this is & why it challenges the status quo of single cell analysis🧵🧵

Narges Razaviannarges@sigmoid.social
2023-05-04

FDA better catch up because these tools seem actually quite useful and proper evaluation frameworks like these may help understand when to be cautious...
---
RT @BoWang87
Huge congratulations to three of my stellar PHD students: John Giorgi @johnmgiorgi, Augustin Toma @ugustintoma, and Ronald Xie @RonaldXie1 ! Our team won first place in the MEDIQA-Chat challenge of clinical note generation from patient-doc conversations hosted by #ACL20
twitter.com/BoWang87/status/16

Narges Razaviannarges@sigmoid.social
2023-05-04

RT @johnmgiorgi
Our preprint summarizing our first-place submission to the MEDIQA-Chat shared task for automatic clinical note generation from patient-doctor conversations is now up on arXiv! Some interesting findings below 🧵👇

arxiv.org/abs/2305.02220

Narges Razaviannarges@sigmoid.social
2023-05-03

RT @AsmaBenAbacha
Our @eacl Paper & Presentation on "Clinical Note Generation from Doctor-Patient Encounters" are now available, Paper: aclanthology.org/2023.eacl-mai
Presentation: youtube.com/watch?v=GUFwiqQO69
Authors: Asma Ben Abacha, Wen-wai Yim, Yadan Fan, and Thomas Lin. @Microsoft & @nuance. #EACL2023

Narges Razaviannarges@sigmoid.social
2023-05-02

RT @PierreEliasMD
Are LLMs able to accurately summarize medical evidence? Is there a difference between #GPT3 and #GPT4? Our preprint, led by @LiyanTang4, had MDs review the accuracy. See results below! @ColumbiaDBMI @UTCompSci @ColumbiaMed @OpenAI twitter.com/liyantang4/status/

Narges Razaviannarges@sigmoid.social
2023-04-28

RT @NIHDataScience
Did you know @AIM_AHEAD has several open funding and fellowship opportunities? The program establishes partnerships to enhance the participation and representation of researchers and communities currently underrepresented in the development of #AI/ML models. #NIHData 🧵

Narges Razaviannarges@sigmoid.social
2023-04-27

RT @oanacamb
📢 We are very excited to have Lisa Hendricks from @DeepMind in person at our @ucl_nlp Meetup on the 16th of May at 5pm BST to talk about Rigorous Benchmarking and Mitigation of Harm in Language Models. If you are not in London, join us via Zoom (link later)! Sign up below ⬇️

Narges Razaviannarges@sigmoid.social
2023-04-27

RT @DrQiaoJin
🏥 You can directly download 167k patient summaries (PMC-Patients) from @huggingface now!
- 📚 Extracted from case reports
- 🗒️ Summary = admission + labs + diagnosis + treatment + discharge + follow-up notes
- 👥 Annotated with 293k similar patients and 3.1M relevant articles

Narges Razaviannarges@sigmoid.social
2023-04-26

For images check the fingers, and teeth and small things that have fixed count. For videos, lol, check out the chewing and touching.

The future is still in (near) future... =)
---
RT @0xgaut
AI is now indistinguishable from reality.

It's hard to believe, but this ad was AI generated. It's not real.

The future is here.
twitter.com/0xgaut/status/1650

Narges Razaviannarges@sigmoid.social
2023-04-25

In this already low quality diagnosis, the situation is worse for females sand for Black patients.

Female & Racial minorities are at higher risk for dementia to begin with, and are under-represented in "research" data that we look at.

Reality of disparities are much worse.

3/5

Narges Razaviannarges@sigmoid.social
2023-04-25

This work was done by my amazing MSc student (now a PhD student at UMass) @HuiWei15 and our awesome collaborator Arjun V Masurkar.

Great Job, Hui!

Narges Razaviannarges@sigmoid.social
2023-04-25

Overall, this is what's happening in a "Research" (i.e. higher quality) cohort. In the actual clinic, the diagnosis qualities are likely a lot worse.

Do AD-specific clinical trials work for LBD? (i.e. Amyloid beta vs lewy bodies vs Tau vs ... )?
Can we do better with AI/ML?

4/5

Narges Razaviannarges@sigmoid.social
2023-04-25

Lewy bodies mostly manifest as mixed AD + LBD (N=140 vs N=572) but in the clinic most of those folks are diagnosed with AD only. Lewy body diagnosis accuracies are also poor across all stages (<47%). This negatively impacts lewy body clinical trials (and success).

2/5

Narges Razaviannarges@sigmoid.social
2023-04-25

We have a paper on accuracy of dementia subtype diagnosis (Alzh vs Lewy body) by clinicians (vs autopsy).

The accuracy is terrible.

Among those diagnosed with AD in clinic, 32% have Lewy bodies (undiagnosed). Mixed AD+LBD sensitivity is just 3%.

frontiersin.org/articles/10.33

1/5

Narges Razaviannarges@sigmoid.social
2023-04-25

RT @james_y_zou
Thank you @StanfordHAI for a nice article discussing our recent #EchoNet #AI randomized clinical trial!

Original @Nature paper: rdcu.be/c9eZC
@Stanford article: hai.stanford.edu/news/proving-

Narges Razaviannarges@sigmoid.social
2023-04-24

RT @causalclaudia
Thrilled to share that the "Spurious Correlations, Invariance, and Stability" workshop is returning to @ICML2023 this year!

Submit your work (by May 23) on all things spurious correlations! See full CFP in

sites.google.com/view/scis-wor

Narges Razaviannarges@sigmoid.social
2023-04-21

RT @srush_nlp
NYCers. Come over to Cornell Tech for the generative AI hackathon on Sunday. twitter.com/davederiso/status/

Narges Razaviannarges@sigmoid.social
2023-04-21

RT @DrJimFan
How to build *TruthGPT*? I listened to a talk by the legendary @johnschulman2. It's densely packed with lots of deep insight. Key takeaways:

- Supervised finetuning (or behavior cloning) makes the model prone to hallucination, while RL mitigates it.
- NLP is far from done!

1/🧵

Client Info

Server: https://mastodon.social
Version: 2025.04
Repository: https://github.com/cyevgeniy/lmst