Morning NMC NPC AI PG e logged journal club on Clinical reasoning initiated by Eric Topol

UDLCO Summary: A journal club lead from Eric Topol on linked in along with a deep query was posted by BG in the NPC, AI healthcare community following which after some meandering around how to obtain full text and what was the journal study design and what were the potential flaws with the data representation, the original question remained unanswered. 

Keyword glossary : http://userdrivenhealthcare.blogspot.com/2023/11/glossary-of-user-driven-healthcare.html?m=1

UDLC transcripts :

[4/2, 4:13 PM] Bharat Gera: https://www.linkedin.com/posts/eric-topol-md-b83a7317_clinical-reasoning-of-a-generative-ai-model-activity-7180587010833694722-fFbe?utm_source=share&utm_medium=member_ios

[4/2, 4:31 PM] AI AIIMS: If anyone has the paper, please share

[4/2, 4:35 PM] : Unfortunately, Sci-Hub doesn't have the requested document:

10.1001/jamainternmed.2024.0295 🙃😆

[4/2, 9:19 PM] Bharat Gera: Thanks for sharing the document from a trusted source...not able to understand why incorrect clinical reasoning was lowest for residents and highest for LLM..any thoughts?

[4/2, 9:20 PM] Rakesh Biswas: Check out this paper👇

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8858363/

[4/3, 12:03 AM] +44: On a completely different note…..May or may not be relevant here! 

One thing we medics in India don’t get free or cheap access to is Journal articles. 

For eg the lancet publishes something today but we have to go begging around for access! 

I’m sure many docs here face the same! 

Scihub was great but doesn’t work any more!

We’d be happy to pay a small price to a platform that allows us access to these! 

Are there any in India already??

[4/3, 7:01 AM] SP  AI: Found this on a thread :

https://get.deepdyve.com/enterprise-plans/

They have individual plans at $59/month or annual at $499 . Also have a two week free trial . 

Seems to have a broad range of journals

[4/3, 7:41 AM] +44: Id pay this I think. Thank u

[4/3, 8:03 AM] Rakesh Biswas: I wouldn't. 

I would rather share my interest in the article in fora such as these and there's always a chance that someone would be intrigued and interested in the same and would fish it out as illustrated by @⁨Bharat Gera⁩ above with @AI⁩ taking it a step ahead by paying it forward and completing the feedback learning loop here which incidentally was captured and woven back into another loop by Bharat in asking why there's some apparent problem with the math in the paper illustrated in it's table 1. 

While we took some time to understand the math and pulled a key reference number 6, thankfully available full text in PMC, it's still difficult to understand the math and hence I reshare some data from the table for the data scientists here presumably @⁨~Kaustav Talapatra⁩ @⁨~Samar⁩ :

While the total respondents are 232, the number of respondents when you total correct and incorrect comes to 245!

Obviously if both celeb chatbot and attendings get error counts of 11 and 10 and residents 2 and the paper concludes that residents (including fellows) are less than the chatbot and attendings there must be something fishy! 

A hint to the answer probably lies at the fine print at the bottom of the table @⁨Bharat Gera⁩?