Slashdot Mirror
Machine Learning Used To Predict Military Suicides
HughPickens.com (3830033) writes David Wagner writes that a predictive computer model using machine learning methods is helping to identify soldiers in the United States Army most likely to commit suicide. Computers combed through data on more than 40,000 soldiers who'd been hospitalized for mental health problems looking at 421 variables on each soldier drawn from 38 military data systems. Using a method known as "machine learning," the researchers identified roughly two dozen factors that are most important in predicting soldiers most likely to commit suicide. The soldiers most likely to take their own lives were men with past suicidal behavior and a history of psychiatric disorders and criminal offenses, including weapons possession and verbal assaults. Soldiers with hearing loss also faced heightened risk — a strong indicator that they had suffered a head injury. So did enlisting in the Army after age 27, most likely because those soldiers had already experienced trouble finding their way in life. "There's this group that comes to the Army later in life — they're smart, they have skills, they tend not to be married and they have no career or have left a career to join," Dr. Kessler said. "We don't know why they should be at higher risk, but they appear to be."
Murray Stein, co-author of the new study, found that among soldiers recently discharged from psychiatric hospitals, more than half of suicides were committed by just five percent of patients. "The most impressive thing is that they identified this high-risk group in the hospital, and by just focusing on one in 20 of them, you're really dramatically improving your ability to predict," says Dr. Mark Olfson, a professor of psychiatry at Columbia University who was not involved in the study. "Clinicians don't do a very good job predicting suicide risk, even though we think we do."
Murray Stein, co-author of the new study, found that among soldiers recently discharged from psychiatric hospitals, more than half of suicides were committed by just five percent of patients. "The most impressive thing is that they identified this high-risk group in the hospital, and by just focusing on one in 20 of them, you're really dramatically improving your ability to predict," says Dr. Mark Olfson, a professor of psychiatry at Columbia University who was not involved in the study. "Clinicians don't do a very good job predicting suicide risk, even though we think we do."
If "more than half of suicides were committed by just five percent of patients", then all the suicides were committed by just ten percent of patients.
No, that is not what they are saying. What they are saying is that if they screen for suicide likelihood, they can identify a sub-group consisting of 5% of the patients that commits 50% of the suicides. People in this subgroup are about 20 times more likely to commit suicide than the average patient. So it may make sense to focus intervention therapy on that subgroup.
It is also such a general term as to be useless to anyone with some AI know how. Did they perform Bayesian analysis? A neural net? Get into the details. Anyway, I think the general public would be shocked to see how utterly simple some of these algorithms are. Most are just really basic probabilities generated off some input...
FTFA ( http://archpsyc.jamanetwork.co... ):
"Administrative data [...] were used to predict suicides [...] using machine learning methods (regression trees and penalized regressions) [...]."
So, decision trees: http://en.wikipedia.org/wiki/D...