Fernando Zampieri @fgzamp.bsky.social · 04/07/2026A legendary performance. Such an amazing team. I will never forget this match. 120
Fernando Zampieri @fgzamp.bsky.social · 30/06/2026Looking for the key to a door you have long forgotten. 000
Fernando Zampieri @fgzamp.bsky.social · 12/06/2026Counterpoint: Have the simplest type of fluid possible on the shelf that is capable of diluting everything and is as cheap as possible. 030
Fernando Zampieri @fgzamp.bsky.social · 12/06/2026Wholeheartedly agreed. Lack of any effect is the more plausible reason, although the perfectly-balanced -opposing-effects hypothesis is getting traction. I remain attached to the null means null hypothesis. 010
Fernando Zampieri @fgzamp.bsky.social · 11/06/2026My less visceral take: - The gradient was likely small to justify any long term outcomes. - But the gradient apparently was enough to avoid pressors, which is still a relevant endpoint. - I disagree a bit with some theoretical aspects of the trial, but I can see its worth clearly =) 130
Fernando Zampieri @fgzamp.bsky.social · 11/06/2026I deleted because I understood a bit better so will post here =) 130
Fernando Zampieri @fgzamp.bsky.social · 11/06/2026Unsurprisingly, when there is no gradient results are the same. 010
Fernando Zampieri @fgzamp.bsky.social · 10/06/2026The assumption that neutral trials in critical care are a peculiar combination of perfectly balanced groups that neutralize themselves is, most if not all of the time, a fairy tale. 120
Fernando Zampieri @fgzamp.bsky.social · 09/06/2026Book titles everywhere: "Neutral Tones in Interior Design" Book titles in North America: "Beige: The Color That Conquered the Suburbs, Destroyed Joy, and United a Fractured Republic" 011
Fernando Zampieri @fgzamp.bsky.social · 08/06/2026Agreed, but it is time to be transparent that most public health interventions for obesity failed. Specifically where I live, cities are designed for cars, and food became so expensive that the concept that someone can walk/bike to work and eat something healthy became ties to wealth. 110
Fernando Zampieri @fgzamp.bsky.social · 08/06/2026I feel it is unfair to patients to tie this to "lifestyle changes". Only those who lived with obesity (as I did) know how hard becoming active is when you are way above your ideal weight. 110
Fernando Zampieri @fgzamp.bsky.social · 08/06/2026Accessibility seems to be the higher problem now. In Brazil incorporation of the drug to the essential medications basket seems inevitable especially now that the patent has expired. 110
Fernando Zampieri @fgzamp.bsky.social · 06/06/2026Canadian Vigour Center (www.thecvc.ca) officially launched the #HEART platform trial, a platform trial for patients living with heart failure. First pilot trial on ketone supplementation to start soon. A fellowship position is also posted. www.heartplatform.ca 211
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026Literature is warming up and extending e-values to platform trials seems the logical next step. This will allow continuous monitoring of clinical trials while controlling type I error. arxiv.org/abs/2602.06379 arxiv.org/abs/2512.04366 arxiv.org/abs/2606.00878arxiv.orgE-values for Adaptive Clinical Trials: Anytime-Valid Monitoring in PracticeAdaptive clinical trials rely on interim analyses, flexible stopping, and data-dependent design modifications that complicate statistical guarantees when fixed-horizon test statistics are repeatedly i... 010
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026An e-value is the payoff of a $1 bet against the null. Formally: E ≥ 0 with E[E] ≤ 1 if the null is true. Big E = the null lost money = evidence against it. Bonus: e-values multiply across studies and stay valid even if you peek and stop early. The p-value can't do that. 110
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026The overall concept is simple. In a fair betting game, knowing previous outcomes shouldn't allow you to make better bets in future draws. If it does, the null hypothesis is false. arxiv.org/abs/2410.23614arxiv.orgHypothesis testing with e-valuesThis book is written to offer a humble, but unified, treatment of e-values in hypothesis testing. It is organized into three parts: Fundamental Concepts, Core Ideas, and Advanced Topics. The first par... 110
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026Sequential-anytime valid inference refers to methods that guarantee anytime validity and type I error control even with infinite looks. It is entrenched in Markov's and Ville's inequalities and comes from game theory. Major companies use it (Netflix included). 110
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026I have few doubts that sequential-anytime valid inference (SAVI) and e-values will reach clinical trials space soon. What is SAVI and what are e-values? And why should you care? 231
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026Dual LLMs is the way forward to assist research. www.github.com/fzampier/DESAL has a SR I'm doing with LLM friends. First paper using framework (but 100% human checked) was this pubmed.ncbi.nlm.nih.gov/42121165/.github.comGitHub - fzampier/desal: SR/MA on hypertonic saline for heart failureSR/MA on hypertonic saline for heart failure. Contribute to fzampier/desal development by creating an account on GitHub. 112
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026I became a heavy user. Automated most repetitive tasks. We have a pilot study using esophageal balloon where the LLMs stress tested the CRF. One LLM built the CRF, the other tested it controlling Chrome/Screen. Didn't replace humans, but by the time I reviewed it was already good. 100
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026Of course you never should let LLMs see real patient data, confidential files, etc, etc, etc, etc. 010
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026Most LLM models can also build a CRF from a draft (or even the protocol) and create a CSV you can simply plug in RedCap. Then you can test it with coaching, insert test data, download, analyse, etc. 320
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026You can let some LLMs control your Google Chrome, including clicking and viewing your screen. This is an easy set up with Claude. You can open RedCap and see the LLM showing you where to click and ask for coaching which it will gladly do. 110
Fernando Zampieri @fgzamp.bsky.social · 04/06/2026Building RedCap is like learning a trade. Best way by far is to learn by doing using LLM of choice + Chrome MCP on a training project. 100
Fernando Zampieri @fgzamp.bsky.social · 03/06/2026Is there an equivalent to the old-forgotten tweetorials in the fancy place here? Is the concept still a thing or am I already the old guy? 120
Fernando Zampieri @fgzamp.bsky.social · 02/06/2026Clustering and phenotyping was largely not a good idea. I'm guilty of this sin. 020
Fernando Zampieri @fgzamp.bsky.social · 22/06/2025We will probably reach interplanetary life before pivoting a data.table becomes an easy and logical task. 010
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025Also depends on time of the day and how busy servers are. You can get a great response followed by an awful one depending on how busy servers are and, of course, if you are using free or paid versions. 000
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025Finally, acknowledge LLMs helped you writing the analysis code in your paper. 010
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025Once your analysis look good, you should: 1. Make sure you understand what's happening. 2. Test, test, test. 3. Ask for a critical review from one or two LLM. Go back if needed. Confirm with a statistician. 120
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025b. Go for another LLM. Gemini 2.5 is good, and can also write artifacts (called Canvas). You, of course, need to provide all context to it. Just pasting your R code and saying "Fix this" won't help, unless it is a simple issue like missing brackets. 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025a. Open a new conversation and project and supply the broken code. Provide evidence on where it seems to be broken. Ask for a fix. This may help (+) 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/20255. Using a single LLM. While Claude is great, sometimes it derails and becomes blind to its mistakes. The first thing you must try to do is fix it yourself, but sometimes it's hard. You can proceed with two approaches: (+) 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/20254. Not create a writing style. This can be done easily in Claude. You can even add papers you have written that are open access so Claude knows what you are usually talking about. (+) 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025I only move to vibe coding when I am sure Claude understood what I am trying to do and I understood what they did on first pass. They need to grasp the idea of the project and my coding style and I need to understand what's happening. 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/20253. Failing to provide context. When I start coding with Claude I usually create a project, update the code, and the first thing I do is asking Claude if he understands what I am trying to accomplish with that code. Then, if they do, I proceed asking them to clean it a bit. (+) 210
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025This avoids a lot of "That's fantastic!", or "Good catch!", etc. 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025For example, my default comments for Claude include: "Do not flatter me. If I am wrong, say I am wrong. If I try to correct you and you still feel you are right, debate. Do not make compliments. Provide direct answers whenever possible" (+) 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/20252. Failing to tweak the LLM to taste. In Claude, you can create projects where you can add files for context, write what you plan to achieve and how Claude should behave. In the overall Claude settings you can also provide an overall guide to how you want the replies (+) 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025...Code should use base R and {marginaleffects}. Code should be as succinct as possible. Please confirm you understood before proceeding. Also please make sure the code uses dead as the outcome, and not as reference". This will produce a much more useful output. (+) 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025...I need to run a logistic regression and extract the effect in OR and RD. In addition, I need also an adjusted model for covariates X, Y, and Z, all continuous, and then get the marginal effects for the model. I finally need a plot comparing the effects from adjusted and unadjusted models.... 110
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025Consider: "Hi Claude, we will be working with R today and I need artifact for a code I plan to run. I have a data frame loaded with outcome labelled as outcome (binary, alive or dead) and intervention (binary, 0 for control and 1 for intervention)... 100
Fernando Zampieri @fgzamp.bsky.social · 16/06/20251. Believe they are talking to Alexa or Siri. This is the most common one. Many people make short, small prompts as if they want Alexa to play a song. For example, people will go and ask Claude the code for a glm. Claude will give back a generic code, if anything. (+) 100
Fernando Zampieri @fgzamp.bsky.social · 16/06/2025I've been using LLM for a bunch of stuff, specially coding with R. Claude is great. I believe some people fail to see their true value because they make simple mistakes: (+) 252
Fernando Zampieri @fgzamp.bsky.social · 15/06/2025Agreed. I guess we didn't try hard enough to further validate the concept. The last 10 years have failed to give the nextlgical step forward: Validate prospectively inside a trial. The longer it takes, the more skeptical people will be. Adding more attempts to the mix is not helpful, I guess.... 110
Fernando Zampieri @fgzamp.bsky.social · 15/06/2025We need less phenotyping papers in critical care. Time of exploring the method is long gone. Time to validate. 031
Fernando Zampieri @fgzamp.bsky.social · 19/02/2025Discovering evidential (likelihoodist) statistics was probably the coolest academic thing for me over the past 6 months. 031
Fernando Zampieri @fgzamp.bsky.social · 13/02/2025That seems a lot of legwork to say "more evidence is needed", heh? 110
