Austin Meyer @austingmeyer

@DrDiGiorgio @EvidenceOpen This sort of AI is built into Epic at our institution now. It makes hospital courses, daily insights, chart reviews, give specific citations in the chart and links them to specific notes.

Technology to verify citations is easy and numerous tools already exist. They just have to be implemented. It would be trivial to have one or several AI models do initial screening for citations (both meta-data and appropriateness) and whatever else was high yield. Not to mention, we could just require doi’s for every citation and have an editorial assistant click on them. That would take 2 minutes.

See the pre-print of our new article. Quantifying the Protection Gap: RSV Activity Outside the Recommended RSV Prophylaxis Administration Window academic.oup.com/jpids/article-…

One additional thing to keep in mind, is the RLHF that general models undergo for alignment is likely to make real time unintentional biasing/steering a significant problem for non-crystallized case presentations. There would probably need to be some specific alignment tuning to this use case.

This is a nice thread. I feel like a lot of non-physicians do not understand the caveats here. I might go farther and argue that I'm not sure it is actually ready for prospective trials except perhaps in a very limited sense. Ultimately, the training data includes lots of heavily curated case presentations from the published literature and essentially no real world presentations with confusing and disorganized thoughts/data. The biggest sticking point is likely to be that very large models should be able to memorize or at least easily pattern match test data that was very similar to the curated content in the training data, but whether they are useful at all on a stream of consciousness presentation (whether presented by physician or a patient) is unclear. I'm not sure it is even clear that a model can consistently take in the information from a patient and present it as a completely hallucination free curated case.

Gemini 3.1 gets citations wrong frequently. Even if it doesn’t hallucinate the title of the article, it hallucinates the other metadata of the citation or inserts the citation for something that the citation doesn’t support. The GPT and Claude models are, admittedly, much less likely to hallucinate. Given the highly variable expertise of authors, it might make sense to still have them note what models were used. I agree that AI review should be at least allowed (many journals strictly prohibit it). It would be particularly useful in medicine where most authors are subject matter experts and relatively poor with methods. Though it is worth mentioning that AI reviews often bring up lots of relatively meaningless pedantic (non-)issues. So reviewers with subject expertise are still needed to strictly prioritize or eliminate concerns raised by AI. Models also tend to lack nuanced subject matter intuition. Since most academic papers are being written by subject matter experts, the main value of AI reviews is probably in methods verification and replication rather than discussion of implications.

I agree that it is definitely a good thing for things to be cheaper. It’s just that the thread was about saving lives. I worry about the narrative that extensive imaging saves lives because it can be surprisingly insidious and dangerous. Especially coming from important and influential people.

Isn’t this a bit too strong? GRADE clearly has issues, but saying that only prospective controlled trials should count as evidence seems like a kind of epistemic fundamentalism. There has to be a path for accepting other types of evidence when prospective controlled studies are infeasible, unethical, or not yet complete. There are lots of decisions physicians have to make with limited evidence and there needs to be a pathway for physicians practicing away from major centers to know what most of their colleagues would do in a certain situation. Do we just have a different outlet for that information?

To be clear, there is no consistent evidence that asymptomatic/average risk people have improved outcomes from full body MRI, regardless of cost or time. It seems intuitive, but there are a lot of relatively risky invasive tests that come after an incidental finding. In many cases, it would be better not to find them while randomly imaging asymptomatic people.

Does it really explain a 60% drop though? It just seems quite unlikely to have such an obvious alternative explanation coincidentally occur, and honestly looking at fairly imperfect plots feels like reading tea leaves. There is also a national decline of something like 10% from peak, but one would expect programs near Silicon Valley to be more sensitive to this.

It may also be that a portion of eventual CS grads were transfers in and those have complete stopped? Or that there was some incentive to complete dual majors with another science or engineering and those have stopped? It just seems like the other models to explain this massive drop are considerably less powerful.

I don’t understand why this can’t be the AI narrative. Most students have the flexibility to switch majors through the first two years of their program. This plots suggests exactly that. Per the chart, the incoming students could be naive and making decisions based on parent preferences. Once they get to campus, they have two years to switch majors before they get locked in. They hear about job market concerns and switch before year three. That would be a perfect timeline given the release of GPT-4 in May 2023.

Our Method/ology paper. An Introduction to Machine Learning for the Pediatric Hospitalist publications.aap.org/hospitalpediat…

Most importantly and probably the biggest bottleneck, is they need to stop testing against benchmarks that consist of cases that have been well curated and summarized by medical professionals. If people want to see how well these models perform, they need to allow random patients to input their own queries and allow the models to ask their own questions in response. The cognitive disorganization of real patients without a doctor filtering it for the models is probably a significant limitation in training.

@mcuban The level of hubris in the tech community is remarkable.

We submitted a paper to a prominent Pediatrics journal that found, "... our analyses suggest that a fixed window would need to be extended to be both a month earlier and a month later (September–April) to more definitively capture the entire RSV season." It was rejected 6 weeks ago (and is under review at a different journal) because the reviewers felt like an additional season of data was required to be sure. Here is the additional season from @AAPNews. publications.aap.org/aapnews/news/3…

@BradSpellberg Not to mention there are so many health systems pushing to remove procal from laboratory availability, which seems crazy to me.

I'm not sure one has to require a virus be present for non-severe, right? Just that bacteria not be present. There are so many fully negative panels and most standard sputum arrays that I've seen don't have any typical bacteria on the array. To be safe, one could define a set of trials sequentially reducing duration with a small set of antibiotic options. Then, test non-inferior reduction from 5 days to 3 days. Then, 3 days to 1 day. Then, 1 day to 0 zero days.

Yes. It just seems like all of the evidence in inapplicable to the most common CAP prescribing context. As you noted, I can't find any actual evidence that we should be giving empiric antibiotics in the outpatient setting for microbiologically typical CAP coverage based on clinical or radiographic evidence for CAP in developed countries (those with high pneumococcal vaccination rates)? As far as I have been able to find, virtually every trial is a non-inferiority trial based on prior standard dating back to penicillin use in case series of severe pna in the inpatient setting prior to pneumoccocal vaccination and not in mild or moderate disease. Almost all non-inferiority trials show that ever increasingly short courses of antibiotics are non-inferior. It seems like we should really go back and study this in the right (modern) context.

Thanks. Are there any for typical pneumonia also? I believe all of those were atypicals. I'm just having trouble finding any for typical CAP in the outpatient setting (where there likely wouldn't be microbiologic confirmation). I'm mostly wondering because most of these radiographic pneumonias are viruses or atypicals.