Still, for many people, it's more trial and error than it needs to be. Many family doctors still don't match up symptoms with neurotransmitters, and then put people on the wrong class of drugs altogether (e.g. something like Wellbutrin which works on norepinephrine and dopamine instead of an SSRI for anxiety which is more associated with serotonin).
Part of the reason is that new drugs keep coming out so it's difficult to keep up, and eventually they all start to blur together.
The best practical predictor of success on a given medication today is if you have a close relative who's had success on that medication.
So include the social graphs of people who are on antidepressant discussion boards and pick up on friends, family, socioeconomic features ...
I so want to use ML tools on medical treatment data, but it's so hard to come by. Have toyed with a couple of startup ideas based on the tendency of people who share a condition to establish ties. Crohn's disease, cutaneous lymphoma. But the groups are too small. Mood disorders on the other hand...
But for some people the results are miraculous. They solve real problems with few if any side effects. For others, they either don't work at all or have intolerable side effects.
The situation would be a whole lot better if we could predict likely efficacy and side effects in advance, especially cases where treatment is more likely to do harm that to help.
Interestingly, I've read a lot (not so much recently) on the subject and taken several SSRIs myself (with good short term experience, minor sides), but I learned a lot from the discussion here. So many people take these drugs — the data necessary to understand and perhaps to predict ... is out there. Not easy to get at, but really plentiful. Reading this made me think about the feasibility of mining message boards for first person accounts of SSRI experience.