Michel Cucherat

OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws In a landmark study, OpenAI researchers reveal that large language models will always produce plausible but false outputs, even with perfect data, due to fundamental statistical and computational limi...

La dernière étude d'OpenAI prouve mathématiquement que les grands modèles linguistiques produiront toujours des résultats plausibles mais faux (hallucinations), même avec des données parfaites, à cause de la nature statistiques des algorithmes en jeu.

www.computerworld.com/article/4059...

Someone was trying to take a portrait of their two fave Chickens, and this happened.

CRISPR is amazing

C'est joli les scores, ça fait des publications, ca donne l'air malin, ca dit que les patients tres malades sont très malades et ont plus de risque de mourir, c'est poussé par des KOL

Et quand on les évalue correctement...

INSPECT-SR: a tool for assessing trustworthiness of randomised controlled trials The integrity of evidence synthesis is threatened by problematic randomised controlled trials (RCTs). These are RCTs where there are serious concerns about the trustworthiness of the data or findings....

www.medrxiv.org/content/10.1...

INSPECT-SR: A tool for assessing trustworthiness of randomised controlled trials.

Alzheimer : polémique autour du Leqembi, un médicament à quelques milliards d’euros

Alzheimer : polémique autour du Leqembi, un médicament à quelques milliards d’euros

Par @olivierhertel.bsky.social

If you are analyzing clinical data, you almost certainty don't need a multivariable model unless you are building a prognostic model or asking causal questions. Statistics shouldn't be a performance.

People often say their goal is to identify “risk factors”. But what does that mean? Some people use the term to indicate potential causes of outcomes. Then just say cause. Others use it to identify predictors of outcomes. Then just say predict. And, sadly, too many others use it as shorthand for factors that are “statistically associated” with outcomes. In this case, say nothing at all, since this “goal” has no clinical utility whatsoever (beyond what it might suggest about causation or prediction). So once you have framed your research question as description, prediction, causation or measurement, there is no longer a need to talk about risk factors. It's basically just a catch-all phrase to cover up muddy thinking.

Relevant:

(from statsepi.substack.com/p/sorry-what... ICYMI)

I'm not saying you can't generate a hypothesis from data.

I'm just saying that generating a hypothesis from the entirety of human knowledge that preceded your data is a much safer bet.

Inclusion of Retracted Studies in Systematic Reviews and Meta-Analyses of Interventions This systematic review and meta-analysis identifies reports of systematic reviews that included retracted studies in their meta-analyses, and assesses the impact of these retracted studies on the resu...

Read our paper in JAMA Internal Med with a great editorial. Wonderful collaboration with @gcabanac.cpesr.fr

Retracted papers are included in meta-analyses published in HIF journal. Inclusion of the retracted studies changed
effect estimates.
jamanetwork.com/journals/jam...

Evidence triangulation in health research - European Journal of Epidemiology For many important questions about influences on clinical and public health outcomes, no single study can provide a decisive answer. The perfect study—a large, diverse, well-conducted trial randomizin...

Triangulation of evidence in medical research explained, in work led by Sirena Gutierrez and @mariaglymour.bsky.social Will do a thread when I get a moment
link.springer.com/article/10.1...

Provenance and funding of extremely cited biomedical papers published in 2003- 2004, 2013-2014 and 2023-2024

Definitions of Validity Terms for Use in Discussions of Randomized Controlled Trials - Journal of Clinical Epidemiology www.jclinepi.com/article/S089...

Analysis of indications for selectively missing results in comparative registry-based studies in medicine: a meta-research study - Research Integrity and Peer Review Background We assess if there are indications that results of registry-based studies comparing the effectiveness of interventions might be selectively missing depending on the statistical significance...

Analysis of indications for selectively missing results in comparative registry-based studies in medicine: a meta-research study

Image shows website.

Viewpoint: Proposals to lower or eliminate the @fda.gov's preapproval effectiveness requirements for drugs are unnecessary and dangerous, as they undermine patient safety and public health. https://ja.ma/4brnk39

‘Arbitrary’ Hodgepodge of Composite Endpoints Used Across ASCVD Trials ‘Arbitrary’ Hodgepodge of Composite Endpoints Used Across ASCVD Trials hdellabella@crf.org Thu, 02/20/2025 - 16:31 Michael O'Riordan A minority of studies used the standard three-point MACE endpoint, and many tweaked their endpoints along the…

ICYMI: ‘Arbitrary’ Hodgepodge of Composite Endpoints Used Across ASCVD Trials

This is sad...

A Comparison of Statistical Methods for Time‐To‐Event Analyses in Randomized Controlled Trials Under Non‐Proportional Hazards While well-established methods for time-to-event data are available when the proportional hazards assumption holds, there is no consensus on the best inferential approach under non-proportional hazar...

doi.org/10.1002/sim....

Our paper "A Comparison of Statistical Methods for Time-To-Event Analyses in Randomized Controlled Trials Under Non-Proportional Hazards" got published today 🎉

We describe commonly used methods, and compare their performance in a simulation study across different scenarios.

Figure. Regression Discontinuities of Nirmatrelvir-Ritonavir Prescription, COVID-19–Related Hospitalization, and All-Cause Hospitalization and Death Rates per 100 000 Older Adults, April 1–November 30, 2022, Ontario

A more than doubled rate of nirmatrelvir-ritonavir prescriptions was not associated with reductions in hospitalizations or mortality among highly vaccinated older adults in Canada.

ja.ma/3ENOh4U

#MedSky

Evidence Lacking for Many AI Tools Used in Primary Care This Medical News article is an interview with study author María Villalobos-Quesada, PhD, a researcher at the National eHealth Living Lab at the Leiden University Medical Centre in the Netherlands.

"We cannot think that our AI systems will not be biased. There will always be a risk of bias."

María Villalobos-Quesada, a researcher at Leiden University Medical Centre, discusses the need for transparent and consistent reporting of quality for #AI tools with JAMA+ AI.

#MedSky

Author’s reply : “The importance of properly specifying your target trial emulation: commentary on Mésidor et al.” - Journal of Clinical Epidemiology www.jclinepi.com/article/S089...

Renowned scientific integrity investigator endows fund to support fellow sleuths Microbiologist Elisabeth Bik donates $200,000 to support training, travel

Many #researchintegrity sleuths are volunteers with little financial support for their often-daunting work to uncover #scientificfraud. One of the most prominent, Elisabeth Bik, has now provided money to change that. @elisabethbik.bsky.social @science.org www.science.org/content/arti...

Uncertainty of risk estimates from clinical prediction models: rationale, challenges, and approaches Clinical prediction models estimate an individual’s risk (probability) of a health related outcome to help guide patient counselling and clinical decision making. Most models provide a single point es...

**NEW BMJ PAPER**

"Uncertainty of risk estimates from clinical prediction models: rationale, challenges, and approaches"

- most models provide just a risk estimate
- we argue for presenting associated uncertainty too
- includes pros, cons, PPIE & methods

Hope helpful!
www.bmj.com/content/388/...

Short news on the RetoRes website inviting to read the new and insightful @jclinepi.bsky.social paper by @davidit.bsky.social and @jdwilko.bsky.social about trustworthiness checks.

Common Sense Oncology principles for the design, analysis, and reporting of phase 3 randomised clinical trials Common Sense Oncology (CSO) prioritises treatments providing meaningful benefits for people with cancer. Here, we describe CSO principles aimed at improving the design, analysis, and reporting of rand...

"Common Sense Oncology principles for the design, analysis, and reporting of phase 3 randomised clinical trials"

Cochrane review: 47% of all clinical trial results are not made public Nearly half of clinical trial results are not made public, an exhaustive review published by the medical evidence group Cochrane has found. The new Cochrane review draws together data from 204 existin...

Interesting 🧐 reading

Using artificial intelligence to semi-automate trustworthiness assessment of randomized controlled trials: A case study - Journal of Clinical Epidemiology www.jclinepi.com/article/S089...

You wait ages and then two arrive at once: reporting guidelines shouldn’t be like buses - Journal of Clinical Epidemiology www.jclinepi.com/article/S089...

Risk of bias assessment tools often addressed items not related to risk of bias and used numerical scores - Journal of Clinical Epidemiology www.jclinepi.com/article/S089...

Uncertainties about the benefit-risk balance of oncology medicines assessed by the European Medicines Agency Drug regulators assess and describe uncertainties regarding treatment outcomes and the benefit-risk balance of newly authorised medicines. We aimed to…

Uncertainties about the benefit-risk balance of oncology medicines assessed by the European Medicines Agency