Very proud of this piece I wrote with Arjun (Raj) Manrai -- a thread will be upcoming after I recover a little from the cognitive load of my week on service, where I *wish* I had an AI second opinion consult service!

Could an AI second opinion service improve medical diagnosis and reduce medical errors **today**? We believe the answer is yes with proper oversight.

Adam Rodman and I reflect on the urgency in our new STAT essay:

statnews.com/2024/05/03/ai-…

Very proud of my BIDMC_Rheumatology colleagues getting involved in the Tweetorial game! If you are interested in rheum and med-ed -- give them a follow! (all credit to the great James Haigney)

A fascinating thread on data storage from the Apollo missions

This leaderboard ranks medical LLM AI models that you can use for #MedEd and clinical tasks such as note summarization, entity recognition, and classification. The top entry currently is an open and free model (OpenBioLLM):

huggingface.co/spaces/openlif…

Oncologists replied to simulated patient messages or edited GPT4 drafts. Edited GPT4 drafts had fewer calls to action, but docs found improved efficiency in >75% of cases. GPT4 impacted physician clinical reasoning, the impact of which MUST be assessed. Great study The Lancet

Since it's #AlienDay , here's a timely reminder:

Top: Sunset on Mars

Bottom: Sunset on LV-426

An impressive study (not quite randomizing, but getting something close in a QI setting) from UCSD:
jamanetwork.com/journals/jaman…

A lot of thoughts about this, but also a general observation -- perhaps we should study these technologies before widely deploying ...

Do EPIC's generative AI replies to patient queries actually save time?

No.

17.9% INCREASED reply length
21.8% INCREASED read time
-5.9% DECREASED reply time (non-significant)

...'not anywhere near where it needs to be to be useful'

jamanetwork.com/journals/jaman…

Amazing that a chip bricked in 46-yr-old Voyager I, preventing it from sending data, and NASA figured out how to split up and reallocate its functions to other hardware, sending code 15 billion miles away (45 hours round trip!)--and Voyager's back online. blogs.nasa.gov/voyager/2024/0…

Fantastic study (and great research methodology) on the abilities of LLMs to perform evidence appraisal. Certainly something a lot of us have been hoping for.

TL;DR: humans outperform LLMs alone, but human + AI performs quite well in some settings.

Randomness is hard to achieve.

It is why the security of 10% of the internet is secured by a wall of lava lamps watched by a camera to generate true randomness

We actually prepared this AM so the chaos that you see at ACP #IM2024 is really just us!

I'll see everyone tomorrow ACP #IM2024 to debate, 'Are SOAP Notes Clean or Dirty?' with Heather Hofmann, MD, FACP and Kristen Toups! It's a follow up to our piece in Journal of Hospital Medicine (pubmed.ncbi.nlm.nih.gov/37530094/)

(I think you can imagine where I stand...)

We are INCREDIBLY excited to host (real magician) Jonathan H Chen MD PhD and BIDMC IM Residency alum Ivana J. to talk about their research in AI in medicine. Anyone at BIDMC IM Residency is welcome -- Deac 312/315 at 6 PM!

(not on Zoom -- in person only!)

Guys -- my 🔥 take is IN SUPPORT of professional self-regulation through the ABIM and other professional societies. What I want is better MOC -- which I think is feasible AND important.

In assessing accuracy for diagnosing curated clinical vignettes from JAMA & NEJM, GPT-4V (which includes image analysis) was superior to GPT-4, GPT-3.5, Gemini Pro, Llama 2, and Med42, & human responses to NEJM subscriber responses jamanetwork.com/journals/jama/… JAMA