Thread by @walterfwiggins

"Kids are not simply little adults."

This adage is repeated countless times throughout med school re: physiology/pharma.

I think it should apply equally to #AI. That's why this headline from @AuntMinnie concerns me greatly.
#RadTwitter #PedsTwitter
1/n

bit.ly/3xH7oX4

In the referenced article, the authors set out to validate @Lunit_AI CXR tool on a cohort of pediatric CXRs.

The authors explicitly state in the abstract that the tool was developed for ADULTS, not kids.
2/n

nature.com/articles/s4159…

Here are a few of the reasons why I'm deeply concerned by this:

#1 - We already have learned that #AI tools can fail unpredictably & inexplicably on data that falls within the distribution of the training data.
3/n

#2 - We also know that performance on data that is out-of-distribution can be wildly unpredictable.

By definition, pediatric data is out-of-distribution for a model trained for adults.
4/n

#3 - Researchers in the field are working hard to identify & mitigate potentially harmful AI biases affecting vulnerable populations.

See recent work from @judywawira @ImonBanerjee6 @DrLaurenOR and others below.
5/n

thelancet.com/journals/landi…

#4 - Children are among our most vulnerable populations.

No silly GIF here or link to any reference. This is a cold, hard, serious fact that stands alone.
6/n

The authors of the study in question pitch validation of AI tools designed for adults on kids as a potential way to address a paucity of AI tools for pediatric populations.

Following the above line of logic, this is a dangerous & slippery slope w/ current state of AI tech.
7/n

FINAL POINT: I'm OK with studying generalizability of AI tools to OOD data, but I STRONGLY disagree w/ the interp that this could address a paucity of AI tools for underserved populations.

@FDADeviceInfo, @AuntMinnie & any other orgs covering this, please consider my plea.
8/8