Rob Gallo, MD

Excited to join @UCSF_DHM_SFGH!

@fperrywilson I like the orangish apples and appley oranges comparison… implies that even though the matched groups look similar they may still be inherently different since you are starting with apples and oranges

Enter propensity score matching! Statistical wizardry that pairs similar patients where one got GLP-1s and one didn't. Like finding orangish apples to compare with appley oranges.

About a year ago, I joked with a friend about Ozempic's miracle benefits. Weight loss! Better hearts! Less drinking! "What about side effects?" he asked. "Maybe after 10 years your eyes fall out," I joked. Well, about that... 🧵

@AdamRodmanMD Thank you Adam!

@robjgallo Congratulations Rob!

Excited to join the JAMA Editorial Fellowship Program!

@kylefbutts @selcukorkmaz @cellis212 Isn’t that essentially how these terms are used in mixed effects modeling? I’ll be honest that I find the terminology of fixed and random effects confusing, especially because there are so many definitions (see the linked Gelman blog in other reply)

@robjgallo @selcukorkmaz @cellis212 it’s kind of wild then to compare “indicators for gender/race/etc.” to “unit-specific (random) effects”, right?

when you see a post like this, you just know it’s gonna be incorrect

@kylefbutts @selcukorkmaz @cellis212 statmodeling.stat.columbia.edu/2005/01/25/why…

@kylefbutts @selcukorkmaz @cellis212 My experience is that outside of Econ when people say fixed effects they are referring to group indicator variables. More than anyone being correct/incorrect, it seems use of these terms are just different across fields.

@kidney_boy It may actually be true that speed of correction doesn’t matter for ODS since it’s a low evidence area, but the likelihood that it significantly affects in-hospital mortality is very low in my opinion

@kidney_boy To me this makes it clear this is probably confounding… why would correcting Na faster or slower change mortality? It would have to be one of the most effective interventions in medicine to have that large an effect!

Honey, stop what you're doing, new hyponatremia research just dropped! What's it say? It looks like slow correction is associated with worse outcomes, like death and length of stay! Was it just a small study? No, it was a meta-analysis of almost 12,000 patients! jamanetwork.com/journals/jamai…

@emollick Provocative result we did NOT expect. We fully expected the Doctor + GPT4 arm to do better than Doctor + "conventional" Internet resources. Flies in the face of the Fundamental Theorem of Informatics (Human + Computer is Better than Either Alone).

Our first RCT on using an LLM on diagnostic reasoning is out! And the results are 🔥🌶️... adding ChatGPT did NOT improve diagnostic accuracy or reasoning, and the AI alone outperformed ALL the humans. What does this mean? A 🧵⬇️ x.com/EricTopol/stat…

Only a fraction of the doctors realized they could literally copy-paste in an entire case history into the chatbot and ask it for a surprisingly smart and comprehensive answer. nytimes.com/2024/11/17/hea…

@williamhersh @Squee451 @emollick Also the human comparison is just 4 physicians: 3 from Germany, 1 from US... hard to generalize to say LLM worse than physicians when you are only comparing to 4 physicians. This source of uncertainty also not accounted for in analyses.

@williamhersh @Squee451 @emollick Hard to draw any conclusions from the Hager et al. study. They repeatedly prompted models 20x but treated repeats as independent inflating the sample size from 80 -> 1600... so the analysis is fatally flawed. See link for more info on why an issue. jamanetwork.com/journals/jama/…

Ok. Deleting this and reposting given the Community Note (you can see the original and Note below). The main point doesn’t change in any way, but I want to make sure I am clear in this post that the measurement was diagnostic reasoning & not final diagnoses. A preview of the coming problem of working with AI when it starts to match or exceed human capability: Doctors were given cases to diagnose, with half getting GPT-4 access to help. The control group got 73% score in diagnostic accuracy (a measure of diagnostic reasoning) & the GPT-4 group 77%. No big difference. But GPT-4 alone got 88%. The doctors didn't change their opinions when working with AI. To be clear, this doesn't say AI will always beat doctors - this is a narrow test. It is much more about what this means for the future. As AI models get better, and match or exceed human level performance, what happens? This is one example where it is happening, and we see the issues emerging.

A small randomized trial of generative #AI for diagnosis again (as seen in a few previous studies) shows higher performance for #AI than physicians + AI. May indicate that physicians need to be trained on how to incorporate AI. #google_vignette" target="_blank" rel="nofollow noopener">jamanetwork.com/journals/jaman…

🧐 Is early PEG after stroke a #TWDFNR ? Maybe observing with NG tube is a viable option. @robjgallo and coauthors review the pros and cons. 🔗 …mpublications.onlinelibrary.wiley.com/doi/10.1002/jh… #VisualAbstract by @PatriciaTranMD

World's fastest 🏃‍♀️‍➡️ will be crowned tomorrow. How fast should you send stroke patients for PEG? @robjgallo et al. suggest 🛑 and assess dysphagia risk in this #TWDFNR 🔗 …mpublications.onlinelibrary.wiley.com/doi/10.1002/jh… #VisualAbstract by @PatriciaTranMD

@AnilMakam @gushamilton @syctong Beyond specific individual shortages, it’s worth considering confounding by the underlying reason for the shortages… I don’t know that they are random events isolated from other healthcare system issues or resource constraints

@robjgallo @gushamilton @syctong good point about the limitations of the IV of time of shortage there was an IV PPI shortage also during 2015-2016

@AnilMakam @gushamilton @syctong Also hard to fully control for time-varying confounding given the IV is based on time. Overall, I'm not sure how to reconcile with ACORN.

@AnilMakam @gushamilton @syctong Interesting study, but it assumes that the only change during this time period was the pip/tazo shortage. At first read, that seems like a strong assumption... was the shortage not associated with other drug shortages or changes in hospital resources?