This question seems to be specifically about your idea so I'll answer as such - any general 'can digital activity be used to predict suicide attempts' is probably too broad for this site.
As I see it you want to use what a person types, across all types of digital media, to get some measure of their risk of suicide attempt.
Two key questions in any machine learning problem
- Can we create variables as predictors?
- Is the data there? i.e. do we have labelled data already
As far as creating variables goes a lot of people don't realise that machine learning (or 'AI') isn't about dumping a load of unedited data into an algorithm and letting it figure it out - 90% of the work is in making those variables into something sensible. In this case it may involve a fair bit of psychology. Perhaps you look at changes in message length - has someone gone from writing essays to short answers? (I'm not a psychologist, that may well be way off the mark). The use of particular words or phrases may come into it - you would need to do a fair bit of research into the differences between those who attempted suicide and those who didn't (perhaps ranking words by frequency and see if there are distinct differences).
This brings us to our second question (or perhaps it should be the first) 'is the data there'? You want to predict suicide risk from digital activity so you'll need the digital activity of all different types of people with a label of whether or not they attempted suicide (and when - no point collecting the data two years after their attempt). Do you think this dataset exists? I'm sceptical - the level of detail you would need (both into someone's digital activity and their personal mental state) is unlikely to be recorded.
It isn't to say I think its an idea to be thrown out, only that you need to be realistic about the amount of effort required to do something like this. You would need to carry out research to collect this data, need to plan out your ideas before hand to know what you would like to collect (the worst thing is to find, after a year of data collection, that you didn't monitor a variable you would have now found vital).