The statistics provided near the end of the article are interesting:
The computer was 93 percent accurate -- identifying those with suicidal tendencies over the control group -- while humans were right slightly more than 50 percent of the time with the same groups.
I could see how a computer would be able to move straight to the heart of the matter--looking solely for the triggers it's been trained to catch--which could help improve its accuracy. Human perception is often blinded by too many outside factors. That being said, I am interested in knowing how well the computer performs in continuing trials; I wonder if language is too narrow a scope to identify risk (leaving out other factors, such as history, facial expressions, intonation). The article, of course, does say that the computer is not meant to replace a practitioner/counselor/doctor--it is only meant to aid in diagnosing risk.
Looking at John Pestian's list of publications, I don't see one where he specifically talks about the outcomes of this project--something I'd very much like to see.