@phlegmfighter @drjohnm Depends what’s given and what’s conditional. P (esoph perf | ablation) = low P (had ablation | esoph perf) = not that low. & certainly P (esoph ablation caused the perf | perf & recent ablation) = super high

@reverendofdoubt Yea. K series would make no financial sense for clinical scientists unless there’s an increase in pay cap or FTE limits.

@phlegmfighter exactly! the system-level problem is no correction: - rolled over and did a low-value procedure? should hurt. - inappropriate resistance to a needed procedure? should hurt. worst mal-incentive @ AMCs. In community, repeated interactions matter more.

@phlegmfighter which is to say, it's possible they're in the wrong or it's possible that you're in the wrong. Just the fact that you have to cajole them is not a problem. What matters is who's right (will the patient benefit from them doing the procedure or not?)

@mattstern2 @VPrasadMDMPH An often ignored aspect of modern medical care is that it's *always* placing bets. However, physicians have an obligation to recommend good bets and be clear about the bets we're recommending. We don't even know when nodule surveillance is a good bet.

@motherofswans @VPrasadMDMPH 4 steps for why probably not: 1. incidental = screening w.r.t. this 2. overdx in lung CA common 3. most screening-detected CA is not life saving 4. pt pref distorted re:uncertainty

@phlegmfighter Ernest question - what's the most reliable data you should do anything at all? (haven't read Adam's paper in depth yet, but for the sake of argument - assume they don't meet Lung CA screening criteria)

@AdamRodmanMD @phlegmfighter Interesting! The more I think about it, the more incidental pulm module seems like a perfect use case - clinicians are terrible at it ( generally ) - usually not worth clinicians limited time to dig in to - actually quite nuanced - needle in hay-stack problem

RT @AdamRodmanMD: This is exactly what my co-PI and friend @jonc101x has written about so eloquently in @NEJM_AI in Who's Training Who -- h…

@AdamRodmanMD @phlegmfighter Have to be able to reason abt things like overdiagnosis to get it right Granted, physicians usually don’t - so still juice to squeeze (whether by AI or better training, or both)

@pricklyresearch @nickmmark interesting idea, hard for me to solve for the new equilibrium of things like that (also, main effects would be about incentives). but the idea of tilting the scales more in the public's favor seems right

@reverendofdoubt @nickmmark yea, I think that's one of the unambiguously good things about this AMCs have been a two-tiered system All <faculty> are equal, but some <faculty> are more equal than others.

@drjohnm @elonmusk Exclusion restriction assumption is gonna be a problem. A lot is changing in the research landscape. Should do regular experiments if they're gonna tinker.

@ProfessorPowder Careful- ‘AI detectors’ alone are not accurate enough to support allegations of plagiarism.

@EricTopol Uh.. the improvements in AUC are trivial? And they just omitted the net reclassification index for all the diseases where ECG age didn’t help much? Many more methodologic problems. This is a fatally flawed paper.

@gushamilton Oh my god there are so many methodological errors in that paper I don’t even no where to start But, the funniest is their main result is built on the improvement in dz detection AUC if you add ECG-age… takes 4 sig figs to see.. lol. So bad.

@phlegmfighter @reverendofdoubt @DrToddLee But I agree w you- I’d prefer more critical care trials w surrogate outcomes that are closely related to intervention effects, even if less “patient-centered” We’d know more things with certainty if we stopped shooting for 90d mortality reductions that require 10k person trials.