Co-founder https://teapawt.net/t.co/a9JrgXpEeh - Opensource AI in Medicine CTO https://teapawt.net/t.co/fpagHocOH3 - Enterprise grade on-prem AI for hospital
Joined September 2021
- Tweets1.2K
- Following347
- Followers4K
- Likes276
Pinned Tweet
Today we introduce MedDecider - an open weight family of Jev style decision models for medicine.
MedDecider achieves SOTA results on variety of medical tasks including knowledge questions, ICD codes, trial notes.
9B scores higher than Decider from Perplexity that is 3 times larger.
Today I will be showcasing MedDecider on @AITinkerers in Atlanta.
If you are in ATL join us at 5pm at the Biltmore to learn how to build SOTA medical decision models
I've been testing @OpenBot_ (open alternative to Grok Bot) and I have to admit it's one of the coolest apps I've seen for a while. And it's made in Poland!
I fully moved my bots from grok
- Terra 6.1 as default (as fast and more intelligent than grok)
- Opus 5.5 for advanced bots (research, code)
My biggest paint point right now - lack of native multi account integrations with with Gmail and Drive, grokbot has it, but I suspect it will come.
Knowing when not to answer is the most underrated skill in clinical AI. In a hospital, a model that says "ask a doctor" on hard cases is worth more than one that always answers.
10 ways people are already using Jev-like models in medicine. One of them is in the operating room 🧵
And someone has to keep score. Maziyar runs these models through 669 clinical decisions. When an open one beat Jev outright (643 vs 628), his point was that a hospital can download the weights and run it in-house.
teapawt.net/MaziyarPanahi/status/2…
btw, if you want a medical decision model that runs on your own hardware, we opened ours. MedDecider comes in 4B, 9B, 27B and 31B, gives a probability for every option, and the 31B follows instructions planted in a note 2.7% of the time vs 13.7% for Jev.
huggingface.co/collections/l…
Locally deployed AI came up again and again at SIIM-CAIMI this week. One example from Basia's notes: a system that flags critical findings in pediatric radiology reports, running inside the hospital, checked on 182,000+ reports (99.7% sensitivity, 97.9% specificity).
She also showed Smart Atlas, similar case search over a hospital's own archive. Day 1 notes 👇
teapawt.net/BasiaKlaudel/status/21…
In case you missed the first day of SIIM-CAIMI26, here are my impressions:
@SIIM_Tweets #CAIMI26
Up to 67% of admitted patients have at least one error in their medication history (CMAJ, 2005).
Create your own medication reconciliation checker with MedDecider. It checks the note against what the pharmacy dispensed, drug by drug.
Free Colab notebook: colab.research.google.com/dr…
Great to see our MedDecider-4B taught Darija, the Arabic that Moroccan patients actually speak.
Abdessamad Aabida (@jip7e) did it in 3.8 hours on two free Kaggle T4s. Darija triage went from 47.4% to 58.9%, French medical exams from 60.3% to 81.4%.
This is why we release small open models.
huggingface.co/Jip7e/Hakim-D…
