Ask HN: Can AI detect mask mandates? | Hacker News Reader

Ask HN: Can AI detect mask mandates?

1 pointjjgreen1 comment

I've looked at various graphs of COVID indicators (infections, hospitalisations, deaths) along with the dates of mask mandates and would expect to see some kind of effect of the mandates on these indicators, but I can't personally see anything (even squinting). But perhaps AI can detect such effects, given that it is able (unexpectedly) to detect a subject's race from chest X-rays and other medical imaging [1] and no-one really knows how. Has anyone tried this? Train your model on indicators and mandate dates, test you model by attempting to predict the dates of mandates (if any) using only the indicators?

[1] https://www.vice.com/en/article/wx5ypb/ai-can-guess-your-race-based-on-x-rays-and-researchers-dont-know-how

For the x-ray AI this is possible because a lot of data exist with clear results. One famous model included the patients information in full text on the x-ray images, which the AI quickly picked up on and which was overlooked by humans.

I believe this was with cancer not race. With a good data set you can create a projection with a mathematical tool that has capabilities to predict something. This sometimes called an AI, but includes a variety of methods and recipes in machine learning or adaptive filtering processes.

Mask mandates already are not providing a clear data set. I don't think an AI can help here. There are some studies indicating a 10% decrease in transmission on average, but are those studies are riddled with confounders.

There are various types of masks and different situations. With highly trained personal in full protective gear transmission still reached up to 70% in some settings. If enough viral load is present it is very easy to get infected with the slightest mistake. So this would count for 30 % or more in transmission reduciton, which is better than 0% without masks.

In public, where people are even less capable to wear these mask correctly, this drops below 10 %. Which is basically in the statistical noise for this kind of thing. You will not be able to predict or notice a 10% change in an exponential growing curve that easily.

Especially if it's riddled with confounders and weird onset behaviour.

Reply on news.ycombinator.com