Paul Windisch

Physician Scientist in Radiation Oncology with an interest for software.

[Paul Windisch @windisch_paul]

Just reviewed a paper evaluating 30B-parameter models as decision-support tools for patients with potentially life-threatening conditions. To spell-check the manuscript, the authors used an Anthropic flagship model.

There’s a disconnect here.

[Paul Windisch @windisch_paul]

@pash22 @Sensible__Med @GSCollins @MaartenvSmeden With prediction markets, it's not necessarily about models. Anybody can place bets and they can do this by leveraging all the information they have available

[Paul Windisch @windisch_paul]

Prediction markets are powerful platforms - now people are looking to apply this technology to clinical trials. I discuss the implications in this @Sensible__Med essay.

[Ash Paul @pash22]

Clinical Trial Prediction Markets: Forecasting Evidence or Distorting It?
https://t.co/uFoiDJlLq4 via @windisch_paul @DrMarkLythgoe

[Paul Windisch @windisch_paul]

Since insurance would have to pay for any post-protocol therapies etc. I think this could very well explain a modest OS difference.

[Paul Windisch @windisch_paul]

I suspect it might be due to different patients dropping out immediately after randomization.

[Pierre Blanchard, MD @PBlanchardMD]

No QoL difference with IMPT vs IMRT in oropharyngeal cancer in the TORPEdO 🇬🇧 trial.
How to explain the differences w/ @SJFrankMD 🇺🇸 trial? Planning? Patients? Crossover in the 🇺🇸 trial? Real absence of difference?

Cc @EmmaHall71 @

[Paul Windisch @windisch_paul]

Therefore, proton arm misses 10% of patients with likely lower socioeconomic/insurance status whose insurance wouldn't have covered the treatment.
The IMRT arm misses 10% of patients with the means to simply go elsewhere to get the treatment they wanted.

[Paul Windisch @windisch_paul]

Reviewer used AI to create their review and didn't even bother to delete the "What next?" suggestion. Academic publishing is cooked.

[Paul Windisch @windisch_paul]

3) Also consider this scenario: Trial recruits, good early signals, insiders place positive bets. One guy places massive negative bet to modify odds => patients see odds => trials fails by not recruiting. Guy walks away with $$ even though trial would/could have succeeded.

[Paul Windisch @windisch_paul]

1) It would aggravate the problem though. It's hard to do option trading for very tiny companies and for a big company, a single trial outcome does not move the stock price much. Betting on single trials gives the insiders more power. Plus, some don't require KYC.

[Rorie Simpson @RorieSimpson]

@NicholasZaorsky @NEJM @JAMA_current @endpointarena @Kalshi @Polymarket 1) The insider concern is real and worth taking seriously, but it's important to note these insiders already trade on trial knowledge via the stock market. A prediction market doesn't create that insider problem, it potentially makes it more visible and trackable.

[Paul Windisch @windisch_paul]

2) I fully agree that the current safeguards are imperfect and that is a problem. However, we are not going to solve it by creating incentives for clinicians to not do what's in the best interest of patients and science.

[Paul Windisch @windisch_paul]

Good take! The whole setup essentially funnels early signals (which may be entirely driven by chance) from the trial through insider trading to potential participants.

Good early signal => Fast recruitment
Bad early signal => Trial is dead cause nobody joins

[Nicholas Zaorsky, MD MS @NicholasZaorsky]

Prediction markets for clinical trials sound like forecasting.
But they’re really betting on patient lives.
This introduces financial incentives that can distort science, undermine clinical equipoise, and erode trust.
Medicine should not be a casino.

https://t.co/WyqswLvpkK

[Anish Koka, MD @anish_koka]

https://t.co/WXyKfmyS8e

[Paul Windisch @windisch_paul]

@jksequoia @theliverdoc DO-HEALTH was an RCT but negative for all 6 primary endpoints. Of course, as they measured a lot of secondary outcomes, they found something significant. IMO that shouldn't guide practice though, but rather cause people to run a dedicated trial if they have conviction.

[Paul Windisch @windisch_paul]

@jeremyphoward I tend to agree, but I think another important point is further down in your original thread: Most human researchers won't either. If LLMs can make the ones with potential more efficient, then I'm hopeful we'll at least get more breakthroughs (in part) due to LLMs!

[Paul Windisch @windisch_paul]

@banteg Did you even get rid of ask-questions-if-underspecified?

[Paul Windisch @windisch_paul]

@DrewMoghanaki @CJTsaiMDPhD @JeffBradleyMD @OpenAlex_org Also surprised me when I saw that it ranked just outside the top 10!
I've seen it a few times when creating these rankings that an older trial is ahead of its modern counterpart. I think there's a "If you write a paper on X, you have to cite Y" effect for older publications

[Paul Windisch @windisch_paul]

Top 10 #radonc trials in #lungcancer ranked by 2025 citations:

- Oligomet trials moving up the ranks with CURB (@CJTsaiMDPhD ) in P3
- RTOG 0617 (@JeffBradleyMD ) still in the top spot

Generated with https://t.co/njWj5JspzJ using
@OpenAlex_org

[Paul Windisch @windisch_paul]

Top 10 #radonc trials in #CervicalCancer ranked by 2025 citations:

- Highly dynamic field with all top 3 trials from 2024
- Followed by Sedlis & Peters both of which (trials and criteria) are still frequently seen on board exams

https://t.co/njWj5JspzJ using
@OpenAlex_org

[Paul Windisch @windisch_paul]

Also note that the whole top 10 becomes entirely #MedOnc once you toggle off the "Core Radonc only" filter 😅