When A.I. says you're mentally ill ...
When A.I. says you're mentally ill ...
By Cheryl K. Chumley - The Washington Times - Tuesday,
March 12, 2019
ANALYSIS/OPINION:
The National Institute of Mental Health reports that
nearly one in five U.S. adults suffers from a disorder of the mind, ranging in
severity from mild to nearly incapacitating, and that in 2016, only 19.2
million of the 44.7 million struggling with “any mental illness” at all —
abbreviated in medical lingo as AMI — went for help.
In 2013, meanwhile, America spent almost $188 billion on
mental health and substance abuse disorders, according to the American
Psychological Association.
That means treatments are costly — but if demand were
truly being met, treatments would be even costlier.
So here comes technology offering a solution.
Scientists say it won’t be long before machine-learning
systems take much of the guesswork out of mental diagnoses and eventually even
identify correlations between sicknesses of the mind and physical symptoms of
the body, so as to make labeling and treating faster and easier.
Hair-raising? A bit.
A machine that tells what’s wrong with your mind based on
responses given by your body does seem on the freakish-slash-frightening side
of the emerging technology world.
Still, it’s not all bad. Not all technology in the mental
health world has to be seen as “eee-vul.”
Certainly, chatbots, for instance, could provide some
sufferers a bit of solace. For those with depression, those who are lonely and
even suicidal, those poised for panic attacks — for those and others, the ease
and rapidity with which a quick in-home consultation-slash-session could be
arranged, with the simple flip of an Internet connection, are positives that
are hard to dispute.
Certainly, the old well-worn, oft-overly used phrase, “If
it saves just one —” could apply here.
And certainly, any privately tapped A.I. that guides a
mentally anguished individual back to safety and sanity — when that same
individual might resist going to a therapist’s office and therefore miss out on
the necessary help — well, that’s not just good for that person. It’s also good
for the community and society at-large.
But, the negatives of mind-diagnosing A.I. seem fraught
with pitfalls and dangers, particularly when peering down the path of
unintended consequences.
First off, think of the data collection.
That’s the lifeblood of any A.I. system; if due diligence
and utmost care aren’t exerted, the privacy dings to unsuspecting citizens
could prove disastrous and outrageous.
But second off, think of the potential for over-reliance.
As with any A.I. in the medical sector, it’s one thing
for a physician to use a scan or a test result to bolster human research and
findings. It’s another thing entirely to turn over the testing and researching
and diagnosing to a machine — learned as that machine may be. It’s another thing
to rely so heavily on an A.I. finding that it becomes the main course of action
the medical professional takes. And it’s decidedly different in the world of
physical health compared to the world of mental health.
A patient wrongly diagnosed by A.I. as a high risk for
cancer and subsequently placed on a low-fat diet by the doctor potentially
loses out on some junk food.
But a patient wrongly diagnosed by A.I. as suicidal and
subsequently directed to a treatment facility for further evaluation? That
patient could not only lose precious freedom, but he or she could be labeled
for life with a condition of the mind that, depending on which way the
political pendulum swings, might lead to denials to fly, denials to enter
certain buildings, denials of security credentials, denials to obtain certain
jobs — even denials to purchase or own firearms.
Mental illness can be a stigma that sticks, even when
it’s based on disputed psychology. Just ask a member of the military who
swallows the post-traumatic stress disorder rather than suffer the potential
career stain by seeking treatment.
Fact is, inviting technology to take over, even in small
part, a field that’s so open to human interpretation, that’s so dependent on
comparisons and degrees and contextual analyses, that’s already so vulnerable
to human error, human guesses, just seems a nail-biter of a game.
And, as we may find, sadly too late, it may not be one
that’s even worthwhile to play.
Copyright © 2019 The Washington Times, LLC
Comments
Post a Comment