Research project to help users be more productive and prevent depression

General
#1

This is an application of ActivityWatch data I’ve had for a while but haven’t gotten around to writing down. The idea is twofold:

  • Help the user adopt certain behaviors
    • Such as browsing Facebook/Reddit less
    • One approach is a kind of “focus time” when distractions are blocked (or the user is incentivized to not be lured in by).
    • This is already done by StayFocusd and can also be done with RescueTime Premium (I think).
  • Predict and prevent unwanted mental states (such as depression and anxiety)

The first thing is a lot easier, but the second thing is (to my knowledge) largely unexplored territory (have been speculated about, but haven’t been done).

There is no doubt much value in being able to predict if a person is going down a bad spiral into depression (Scott Alexander has written about this). A large part of the problem has always been how to predict these things with the limited data available when you care about ethics and privacy. That’s where the self-ownership of ActivityWatch data could change things.

The remaining challenge is how to train the predictive model (both how to deal with privacy and how to make it good at predicting something as complex as depression). I hope the research and user community could come together on the data privacy part (some self sacrifice of privacy might be pragmatic).

Let me know what all of you think about this. These forums are pretty dead right now so I don’t expect a reply anytime soon but I’ll link to the thread and update it when progress is made.

1 Like
#2

Linked this to a few friends and they reminded me that the average person probably spends more time in front of their phone and it’d therefore be necessary to log phone activity if you want a sample representative of the average person.

Another good reason to increase the priority of the Android watcher I guess.

#3

Sounds like a good idea overall. A good first step might be to actually document what is known of the field already : if we could know anything and everything about a person ('s activity), how do we predict depression ? Probably having a team member that is actually trained in psychology would be a good idea for that.
And then, how close can we get with AW to the parts that are required to make the prediction. Maybe that will put emphasis on some feature requests like the Android part you detected, maybe it will reveal some other issues, maybe it will JustWork™ and could be effective fast.

You say “train the predictive model” but I don’t think ML will get you anywhere on this, unless you have a good plan for having thousands of people use AW regularly and give you a correct assessment of whether or not they are clinically depressed. Maybe you could run a pilot project to install AW on the devices of all patients of a hospital, and see if from the data you can distinguish the depressed people apart from the others, but I wouldn’t bet on any of that.

(So I’d suggest a predictive system more like hard-coded rules inspired from psychiatrists’ practice, to begin with.)