work with

Diantha Schipaanboord, Floor B.H. van der Zalm, René van Es, Melle Vessies, Rutger R. van de Leur, Klaske R. Siegersma, Pim van der Harst, Hester M. den Ruijter, N. Charlotte Onland-Moret, on behalf of the IMPRESS consortium

Conclusion: The convolutional neural networks in this study demonstrated resilience to simulated sex-imbalance in training ECG data.

pre-print: doi.org/10.1101/2025...

Discrimination remained stable across sexes; only calibration shifted in extreme scenarios when prevalence differed by sex, with similar patterns for women and men.

Using ~165k ECGs, we simulated sex-imbalances in representation (women-to-men ratio), outcome prevalence, and misclassification in the training data for LBBB, long QT syndrome, LVH, and physician-labeled “abnormal” ECGs.

Even if you model a physical system, e.g. avg yearly temperature depending on height, and assume that temp given height is the same everywhere. If you invert it into predicting presence of mountain given temp, you’ll find varying discrimination in diff countries. Example from scholkopf’s talks

You’ve modeled a system with no meaningful variation across environments. The model may be reliable in the tested environments but you haven’t shown robustness against variation in distributions as you haven’t observed any

A question that remains is how these differences in environments may come about and what to do with this in practice? On this, I wrote a paper titled, available here: arxiv.org/abs/2409.01444

fin!

if the distribution of outcome given features remains the same (Y|X), calibration is preserved. If both are the same, the environments were not meaningfully different to begin with!

a more lengthy explanation is in this blog post: wvanamsterdam.com/posts/250425...

as promised (so all of you can breathe normally again), here's my TLDR answer:

Environments must differ with respect to something. If the distribution of features given outcome remains the same (X|Y), discrimination is preserved;

tagging some prediction modelers / statisticians, @maartenvsmeden.bsky.social @benvancalster.bsky.social @gelovennan.bsky.social @f2harrell.bsky.social @lucystats.bsky.social @miguelhernan.org @gscollins.bsky.social

(I will answer tomorrow)

what are the exceptions?

2. an external reproduction of the PROTECT method from Manchester University with Charlie Cuniffe, Matt Sperrin and Gareth Price (www.nature.com/articles/s41...)

3. a 'causal' meta-analysis method using only aggregate data, exciting work with Qingyang Shi from Groningen University

Building in the physics is one way to potentially get the right causal mechanisms

In sofar as the model is trained on real world patient data, you'll still have to ensure no biases e.g. related to confounding creep in

Not sure about overfitting, results seemed robust to 5-site cross validation.

It just learns correlations, what's wrong with that? The words 'confounders' and 'bias' make it sound they expected the model to yield some causal understanding. Maybe these heatmaps are the new table 2 fallacy