Orr Shauly

Plastic surgery chief resident at Emory University. Researcher, device inventor, and weekly Substack writer of The Digital Curbside.

[Orr Shauly @oshauly]

Every time a new benchmark comes out I think the same thing. A model can ace the test and still not help my patient get the right thing done on time. What I want to know is whether it helps outcomes without adding more checking work for the team. https://t.co/eAJdVs9iC6

[Orr Shauly @oshauly]

@YuexingHao @SciReports @MatrAIx2026 Congrats on the paper. Using clinicians and LLM graders over multiple rounds is the part I'd like to see more of. How often did the LLM graders disagree with the clinicians, and which way did they tend to lean?

[Orr Shauly @oshauly]

@weldeiry Nice use of Claude, and I like that you said you checked it before posting. The point about tokenized records still being re-identifiable is the one I think gets the least attention. Who decides when a linkage is allowed still seems pretty unclear to me.

[Orr Shauly @oshauly]

@doctorbhargav Same issue in surgery. A high risk score doesn't tell me whether to operate, wait, or change the plan. I'd like more AI papers to report what clinicians actually did with the score and what happened to those patients.

[Orr Shauly @oshauly]

@EricTopol @uk_biobank @LancetDigitalH Interesting that it comes from routine EHR data and not a special test. What I'd want to know is what a clinician should do differently when someone's ClinicalAge is 10 years older than their real age. Without that it's hard to see how it changes a visit.

[Orr Shauly @oshauly]

When I connected Gmail to Claude I told it to never read patient emails or anything with my bank or insurance info. Took a minute. A physician inbox isn't the place to just click through the permissions. https://t.co/CnENt9y7ah

[Orr Shauly @oshauly]

Raycast keeps my whole clipboard history, not just the last thing I copied. Cmd H pulls it all up. If you copy patient info during the day, clear it out before you leave. https://t.co/6HGo57cPyS

[Orr Shauly @oshauly]

@akiffpremjee Makes sense. Being able to click through to the source note is what would make me trust a summary, even if I only check it some of the time.

[Orr Shauly @oshauly]

@NathanMubasher Curious how they'll define reasonable steps. A lot of us already click accept on suggestions in the EHR all day. Does documenting that I reviewed the AI output count, or will they expect more than that?

[Orr Shauly @oshauly]

@bobvarkey As a surgeon I'm not sure procedural fields are that safe either, since a lot of what we do is deciding who should get the operation. The sign off part is what I keep thinking about. If one clinician signs off on hundreds of plans a day, how careful can that review be?

[Orr Shauly @oshauly]

What made Wispr Flow stick for me was setting push to talk to the Fn key. Hold it, talk, let go, and the text shows up in whatever box I'm in, including most EMR fields.

https://t.co/AEUxzHFEWj

[Orr Shauly @oshauly]

@MaziyarPanahi On-device is the part I like, since a lot of hospitals get stuck on where the audio goes. How are you thinking about the stroke flag? I'd want to know how often it fires on dizzy patients who don't have a stroke before I trusted it in clinic.

[Orr Shauly @oshauly]

@jonathanhershon @Google The 36% figure is the part I'd point health systems to. Admin is the easier pitch, but most of what I'd ask a tool during a clinical day is a clinical question. Curious whether you see different query patterns between residents and attendings.

[Orr Shauly @oshauly]

@nikillinit From the physician side, I'd use something like that if it sits where I already look things up and doesn't need another login. Orders and notes still live in the EHR though, so I'm curious how many of these plug-ins can actually write back to it.

[Orr Shauly @oshauly]

@jselanikio @AnnalsofIM The heavy users writing longer notes is the part I'd look at. When the scribe makes it easy to capture everything, people keep everything, and someone still has to read it. I'd be curious whether inbox time moved at all, since that's where a lot of evenings go.

[Orr Shauly @oshauly]

Try a pre-mortem on your next big plan. Tell the AI it's six months from now and the whole thing failed badly, then ask it why.

I do this for work projects and trips.

https://t.co/4uDjBTfUUD

[Orr Shauly @oshauly]

On Claude Pro and keep hitting your limit? Check which model you're using. Opus burns through it fast.

I use Sonnet for basically everything and save Opus for research and slide decks.

https://t.co/ROjPtdbn4I

[Orr Shauly @oshauly]

@BertWilliamsMD1 Agree. A 15 to 20% drop tells us fewer services got done, not that the right ones were cut. I'd want chart review on a sample of the denied cases before anyone calls it savings.

[Orr Shauly @oshauly]

@zakkohane Nice to see an actual randomized trial instead of another retrospective AUC. The faster time to diagnosis is clear. I'd want to know how many flagged patients got seen sooner and turned out not to have disease, since rheum wait lists are already long.