Maarten van Smeden

Evaluation of performance measures in predictive artificial intelligence models to support medical decisions: overview and guidance Numerous measures have been proposed to illustrate the performance of predictive artificial intelligence (AI) models. Selecting appropriate performance measures is essential for predictive AI models i...

Our guidance regarding performance measures for medical AI models is finally out!

- Stop bashing AUROC, although it does not settle things
- Calibration and clinical utility are key
- Show risk distributions
- Classification statistics (e.g. F1) are improper

www.thelancet.com/journals/lan...

NEW PAPER

The use of explainable AI in healthcare evaluated using the well known Explain, Predict and Describe taxonomy by Galit Shmueli

link.springer.com/article/10.1...

🧐

this is one of my favourite observations about sample size calculations. (afaik first articulated by Miettinen in 1985)

Ha! I did not know I quoted Miettinen :). Thanks for the reference

For some research studies the optimal sample size should be estimated at 0

“Data available upon reasonable request” is academic language for you can get my data OVER MY DEAD BODY

I take version control very seriously

Manuscript_Final_Version_actualFINALcopy_version9b_USETHISONE.docx

Prediction models that are used to guide medical decisions are usually regulated under medical device regulation. This means, putting a calculator out there to promote the use your new prediction model is likely to break some rules.

The lasso works really well in particular settings and for particular purposes. If you are after high prediction performance alone and you have a rather large sample size, it can be an excellent choice indeed. But most analytical goals are not only about prediction

Kind reminder: data driven variable selection (e.g. forward/stepwise/univariable screening) makes things *worse* for most analytical goals

Vacancy — PhD position on AI methodology for prediction of patient outcomes using organoid models Are you passionate about bringing personalized medicine to the next level and make real impact in healthcare? Join our team and develop novel AI methodology to improve predictions of relevant patient ...

NEW FULLY FUNDED PHD POSITION

Looking for a motivated PhD candidate to join our team. Together with Danya Muilwijk, Jeffrey Beekman and I, you will explore opportunities and limitations of AI in the context of organoids

For more info and for applying 👉
www.careersatumcutrecht.com/vacancies/sc...

Interpretable "AI" is just a distraction from safe and useful "AI"

This is right tho. Let’s therefore call them sensitivity positive predictive value curves bsky.app/profile/laur...

Performance evaluation of predictive AI models to support medical decisions: Overview and guidance A myriad of measures to illustrate performance of predictive artificial intelligence (AI) models have been proposed in the literature. Selecting appropriate performance measures is essential for predi...

For details: arxiv.org/abs/2412.10288

No.

I wonder who those people are who come here dying to know what GenAI has done with some prompt you put in

If you think AI is cool, wait until you learn about regression analysis

TL;DR: Explainable AI models often don't do a good job explaining. They can be very useful for description. We should be really careful when using Explainable AI in clinical decision making, and even when judging face validity of AI models

Excellently led by @alcarriero.bsky.social

NEW PREPRINT

Explainable AI refers to an extremely popular group of approaches that aim to open "black box" AI models. But what can we see when we open the black AI box? We use Galit Shmueli's framework (to describe, predict or explain) to evaluate

arxiv.org/abs/2508.05753

Guidelines for Reporting Observational Research in Urology: The Importance of Clear Reference to Causality - PubMed Observational studies often dance around the issue of causality. We propose guidelines to ensure that papers refer to whether or not the study aim is to investigate causality, and suggest language to ...

This is, however, not clever or safe writing, it is a bad collective habit that needs to stop. Not by avoiding references to causality but by clear referencing to it

pubmed.ncbi.nlm.nih.gov/37286459/

The healthcare literature is filled with "risk factors". This word combination makes research findings sound important by implying causality, while avoiding direct claims of having identified causal associations that are easily critiqued.

And taking this analogy one step further: it gives genuine phone repair shops a bad name

When forced to make a choice, my choice will be logistic regression model over linear probability model 103% of the time

Cover picture with blog title & subtitle, and results graph in the background

Post just up: Is multiple imputation making up information?

tldr: no.

Includes a cheeky simulation study to demonstrate the point.
open.substack.com/pub/tpmorris...

The leaky pipe of clinical prediction models. by @maartenvsmeden.bsky.social‬ et al

You can have all the omni-omics data in the world and the bestest algorithms, but eventually a predicted probability is produced & it should be evaluated using well-established methods, and correctly implemented in the context of medical decision making.

statsepi.substack.com/i/140315566/...

Clients: “I want to find real, meaningful clusters”
Me: “I want world peace, which is more likely to happen than what you want”

Depending which methods guru you ask every analytical task is “essentially” a missing data problem, a causal inference problem, a Bayesian problem, a regression problem or a machine learning problem