Hacker News Reader

Comment by ImaCake | Hacker News Reader

A word of caution about machine learning models in academia. The advertised predictive ability or AUC is often inflated, sometimes by a lot. I am not sure why, but I imagine it comes down to the usual p-hacking incentives. Maybe it is because you only need to apply your model to more than one "independent" dataset before you publish a paper. Or simply that academics in diverse fields often lack the nuanced knowledge to realise their model is predicting based on some experimental artifact in their data.

Digital phenotyping has a lot of promise. But considering the unusually conservative old boys club that seems to dominate alzheimers research I would be particularly suspicious of any research on the disease.

View on HN

Fantastic point. Another thing about academic ml models is that they're commonly evaluated on in-sample data. For example, medical AI devices are commonly evaluated on data coming from the same source/hospital that the training data comes from (see https://www-cdn.law.stanford.edu/wp-content/uploads/2021/04/...). This leads to over-confidence in the performance because of overfitting of batch effects.

> But considering the unusually conservative old boys club that seems to dominate alzheimers research

This seems like an interesting claim, worth substantiating? For someone with no knowledge of the field.

Sure, there's a bit of a phenonemon where a small group of labs come to dominate a field of research through a positive feedback loop. They have some initial success, and then can get funding, they then have increased resources to effectively lobby for more funding, squeezing out other groups. Often leading scientists have a lot of influence over the grant funding bodies, but my understanding of these politics is murky. Often this isn't too problematic, but the Alzheimer's field seems really hell bent on the beta-amyloid thing which looks good on the surface but every drug they have tried (and many have been tried) hasn't worked in practice. Here's a nature news article [0] discussing the beta-amyloid failed drugs story. You will note quotes from leading researchers from Harvard and other prestigious institutes still defending drugs against beta-amyloid despite the history of failure.

These leading labs should be looking at different tactics, but the quotes seem to suggest that as of 2018 these groups were still intent on the same strategy despite the failures.

0. https://www.nature.com/articles/d41586-018-05719-4

Reply on news.ycombinator.com