Suicide is notoriously difficult to predict. That meta-analysis is one example; once you account for base rates in realistic scenarios, you can do very little better than chance.
Some of this might have to do with the nature of predictors; maybe with deep learning and real-time data this might improve slightly, but I doubt it somewhat (there are significant ethical issues involved in such data regardless but that's a different issue).
One of the biggest problems in this area is that suicide is such a low base rate phenomenon that it's difficult to predict for that reason alone. Not only that, but many of the predictors are relatively high base rate by comparison, so there's a very very very low signal to noise ratio. Statistically speaking, it's like finding a needle in a needlestack.
I think adding to the complication is that it's not really that you want to know if someone eventually commits suicide, what you want to know is whether they do so within a certain actionable timeframe. That is, for most of the decisions such predictions are being clamored for, it doesn't really help to know that there's a 70% chance that someone will commit suicide eventually; what practitioners really want to know is "what are the odds someone will commit suicide if I make decision X right now." This is a significantly more difficult thing to study.
I personally believe that the approach to suicide in developed countries is horribly misguided and probably unethical, focused too much on the act and not prevention of the state (psychological, sociological, economic) that leads to it. It's unethical to leave someone in such a personal hell that they want to exit it by killing themselves; preventing themselves from doing so, without removing them from that state, only deprives them of a means of relief.
Suicide is notoriously difficult to predict. That meta-analysis is one example; once you account for base rates in realistic scenarios, you can do very little better than chance.
Some of this might have to do with the nature of predictors; maybe with deep learning and real-time data this might improve slightly, but I doubt it somewhat (there are significant ethical issues involved in such data regardless but that's a different issue).
One of the biggest problems in this area is that suicide is such a low base rate phenomenon that it's difficult to predict for that reason alone. Not only that, but many of the predictors are relatively high base rate by comparison, so there's a very very very low signal to noise ratio. Statistically speaking, it's like finding a needle in a needlestack.
I think adding to the complication is that it's not really that you want to know if someone eventually commits suicide, what you want to know is whether they do so within a certain actionable timeframe. That is, for most of the decisions such predictions are being clamored for, it doesn't really help to know that there's a 70% chance that someone will commit suicide eventually; what practitioners really want to know is "what are the odds someone will commit suicide if I make decision X right now." This is a significantly more difficult thing to study.
I personally believe that the approach to suicide in developed countries is horribly misguided and probably unethical, focused too much on the act and not prevention of the state (psychological, sociological, economic) that leads to it. It's unethical to leave someone in such a personal hell that they want to exit it by killing themselves; preventing themselves from doing so, without removing them from that state, only deprives them of a means of relief.