
building personal doctors @nollahealth
New York
Joined July 2011
- Tweets344
- Following528
- Followers618
- Likes4.3K
Pinned Tweet
Nolla Health has received pilot approval in Utah for our AI system to autonomously prescribe acne treatments. To our knowledge, this is the first time a US regulator has allowed an AI to decide whether to start a patient on a prescription treatment.
The pilot is limited to adults with mild to moderate acne and a short list of topical medications. In the first stage, a physician must approve each prescription before it goes to the pharmacy. If the pilot meets its safety requirements, the AI can then send prescriptions first, with physicians reviewing them afterward.
I lead engineering at Nolla, and I’ve spent the last year building this with our team. Incredibly excited for this to launch. This is the a huge step toward understanding treatment decisions AI can safely make, and whether that can help more people access care.
Look, I think the Riemann hypothesis is a wonderful marvel of human ingenuity. but it’s a yetzer harar trap
3
14
602
"The really good doctors we work with are very good at explaining to patients in simple terms what's happening."
From @nollahealth CTO @sean_geiger: "There's a lot of health stuff that feels like it's very complicated."
"But a lot of times the core issues are very simple... Like having a specific imbalance or pathogen here and there."
"(Explaining things simply) helps patients understand how to adhere to treatment paths and how to make personal decisions that help with their health plans the most."
From Industry News with @sophiadew
"We're going to be writing prescriptions fully autonomously with AI."
From @sean_geiger, CTO of @nollahealth: "We can handle entire care paths for treatments, including acne, without a doctor in the loop."
On Nolla's pilot program in Utah: "Utah has set up a regulatory sandbox (for use of AI in healthcare)."
"A lot of other states are considering similar things. The practice of medicine is regulated on a state-by-state basis."
"Utah's been extremely fast-moving... They have an excellent team that's thought about a lot about new capabilities of models come out."
"There are other states that have passed similar legislation to create these sandboxes, but no one has actually gone live with the program yet except Utah."
From Industry News with @sophiadew
About to hop on MTS to talk about our Utah prescription pilot and expanding access to care!
ANTHROPIC IPO | BAN RSI? | TRUMP IN CHINA teapawt.org/i/broadcasts/2WXPOpqnj…
2
25
774
Sean Geiger retweeted
People desperately want better healthcare experiences. Unfortunately, right now, a bunch of tech bros are filling that void with unstudied drugs and unproven diagnostics with compelling interfaces.
The biggest opportunity to improve human health right now may not be the next medical breakthrough. But rather making great medicine as compelling to consumers as the alternative.
Lots of E&M reimbursement will rapidly approach $0 in next 5 years and I’m not sure people are ready for it
1
2
779
Sean Geiger retweeted
Millions of deaths could be prevented with more accessible healthcare and better use of health data.
At @nollahealth, we're building a personal AI doctor for everyone, and we'd love for you to join us.
DM me to join our founding community and get early access to events, opportunities, and the product!
Looking to talk to someone who's brokered healthcare data before. We're looking for some hard-to-find datasets
Purgatory where you’re constantly trying to get to an important meeting at a healthcare conference in a Vegas hotel but you just keep walking in circles and the only people who you run into are all Deloitte employees who are more lost than you
1
10
609
A few interesting things we’ve learned running this benchmark for the last year:
1) Gemini models have always led the pack in single-shot DDX from an image. 2.5 Pro, then the 3.X Flash models, then Argon have all out-performed their Ant/OAI siblings reliably for that task.
You would expect that to translate pretty directly to DDX accuracy in a multi-turn conversation that includes a photo, but here we see much closer performance between the big 3. Anecdotally, OAI has always performed the worst in single shot, but the best in multi-turn, and asks some very clever questions in patient interviews.
2) Not surprising, but: human specialties matter! For DDX accuracy, we see highly specialized, experienced physicians beating the best frontier models when working on cases within their specialty.
E.g. the most experienced Derm in our study handily beat best frontier models.
This benchmark headline comparison includes asking physicians questions outside their specialty.
3) Some models are much better than others at connecting the visual observations with patient reports. The ability for a model to accurately describe an image (e.g. “A well-demarcated, erythematous plaque measuring 1.2 cm × 0.8 cm, with a silvery-white scale on top”) is the most reliable proxy for final diagnostic accuracy, especially in multi-turn tasks. This is also where we still see the host hallucinations in models today.
Today, we’re introducing the Nolla Visual Medicine Benchmark. We’ve been using it internally for the past year to evaluate models for patient care.
Over that year, we’ve seen frontier models make rapid progress. Now we’re sharing the results comparing frontier model performance with human performance.
Today, we’re introducing the Nolla Visual Medicine Benchmark. We’ve been using it internally for the past year to evaluate models for patient care.
Over that year, we’ve seen frontier models make rapid progress. Now we’re sharing the results comparing frontier model performance with human performance.
Our registered observational study included 23 physicians across 10 primary specialties: 11 board-certified physicians, 10 residents and 2 fellows.
They evaluated cases across a broad range of specialties, with most cases outside their own specialty.
1
6
94
If you’re building or evaluating medical AI, we’d love to compare notes.
Here’s the benchmark design, physician study and latest model results:
nollahealth.com/blog/visual-…
1
6
97









