Wednesday, March 27, 2024

AI's not ready for depression diagnoses

The ideas and innovators shaping health care
Mar 27, 2024 View in browser
 
Future Pulse

By Ruth Reader, Erin Schumaker and Carmen Paun

DIAGNOSIS

An iPhone displays the Facebook app.

Researchers found that existing models for detecting depression from text performed poorly for Black participants in a study reviewing Facebook posts. | Jenny Kane, File/AP

How we write our social media posts might reveal whether we’re depressed or anxious. But research methods for finding signs of depression in text don’t work across racial groups, according to a new report funded by the National Institutes of Health. The study is part of an effort by the agency to root out health misinformation among populations that experience higher averages of preventable disease.

Researchers analyzed 868 Facebook posts from people ages 18 to 72 with mild depression. Three-quarters of the posts were by women.

The researchers found that existing models for detecting depression from text performed poorly for Black participants. That remained true even when the algorithms were trained on language exclusively from Black participants.

Previous research has shown that the use of “I” statements and self-loathing or despairing comments can be markers of depression. Those studies and models have been built entirely on phrases and text from white people. Researchers found that the same rules didn’t apply to Black individuals.

For example, depression in Black participants didn’t correlate to self-criticism, self-deprecation, sentiments about being misunderstood or indications of despair or feeling like an outsider — all traditional signs of depression in white individuals.

Why it matters: Artificial intelligence could one day be trained to spot signs of depression in our texts or online posts. But that won’t work if the AI is inappropriately tuned for different populations.

What’s next: The study effectively calls into question current models of detecting depression in language and says new research should look for markers of depression in the text of nonwhite groups. And future research should consider the possibility that depression might not appear in language for Black individuals at all.

WELCOME TO FUTURE PULSE

A photo of a bridge, flowers and a body of water.

Boston, Mass. | Aimee Schumaker

This is where we explore the ideas and innovators shaping health care.

AI is coming for happy hour. Belgian scientists have developed an AI model that can predict how consumers rate certain Belgian beers, according to a new study. Up next: Researchers want to use their findings to improve the taste of non-alcoholic brews.

Share any thoughts, news, tips and feedback with Carmen Paun at cpaun@politico.com, Daniel Payne at dpayne@politico.com, Ruth Reader at rreader@politico.com or Erin Schumaker at eschumaker@politico.com.

Send tips securely through SecureDrop, Signal, Telegram or WhatsApp.  

WASHINGTON WATCH

Dr. Renee Wegrzyn speaks.

ARPA-H Director Renee Wegrzyn has emphasized the need for solutions to make digital health tools reliable and safe. | Josh Reynolds/AP Photo

The Advanced Research Projects Agency for Health is teaming up with the Defense Advanced Research Projects Agency to fight cybercrime.

The two agencies are working together on the Artificial Intelligence Cyber Challenge, a two-year competition for AI and cybersecurity experts to design AI tools to identify and fix software vulnerabilities in critical infrastructure like transportation, emergency services, energy sources and water and wastewater systems.

To help ensure the contest addresses vulnerable health care software, including medical devices, hospital IT and biotech equipment, ARPA-H will commit $20 million toward the challenge.

“In the wake of recent cyberattacks targeting our nation’s digital health infrastructure, there is increased urgency to develop straightforward and dependable solutions to protect patients’ data and access to care,” ARPA-H Director Renee Wegrzyn said in a statement.

“Off-the-shelf software tools aren’t cutting it.”

Why it matters: Health care cyberattacks, especially ransomware and hacking, have surged in recent years, according to the HHS’ Office for Civil Rights.

The OCR reports a 256 percent increase in reports of large breaches involving hacking and a 264 percent increase in ransomware over the past five years.

Last year, hacking made up 79 percent of large breach reports, and in 2023, large breaches affected 134 million people.

Zoom out: A cyberattack last month at Change Healthcare, among the largest U.S. clearinghouses for medical payments, highlights how easily attacks on health care systems can hurt providers, pharmacies and patients.

Health care companies might be especially vulnerable to cyberattacks because they’ve historically invested less in cybersecurity than other sectors, like finance, POLITICO’s Kelly Hooper and Robert King report. When cybercriminals identify an industry weakness, they are relentless — and strike until companies fix the problem.

What's next:Registration for the competition is open through April 30.

AROUND THE NATION

FILE - In this June 3, 2010 file photo, Dr. Steven Birnbaum works with a patient in a CT scanner at Southern New Hampshire Medical Center in Nashua, N.H. New lung cancer screening guidelines from three medical groups recommend annual scans but only for an older group of current or former heavy smokers. The advice applies only to those aged 55 to 74. The risks of screening younger or older smokers or nonsmokers outweigh any benefits, according to the   guidelines published online Sunday, May 20, 2012, in the Journal of the American Medical Association. (AP Photo/Jim Cole, File)

Researchers are testing AI-generated text messages to get people at risk screened for lung cancer. | AP

A sociologist at Washington State University will test how people at risk for lung cancer in a rural area of the state respond to AI-generated text messages encouraging them to visit a local clinic to be screened for the disease.

Two versions of messages will be sent to some 200 patients, one direct and one polite, said Anna Zamora-Kapoor, who’s leading the NIH-funded project that aims to advance health equity and researcher diversity.

Why it matters: The project aims to help rural clinics use the limited resources they have to reach out to a higher number of patients than they’ve been able to in the past.

It should also show which message would most effectively convince people to be screened for cancer.

Zamora-Kapoor’s project targets people between ages 50 and 80 with a history of smoking who would benefit from a low-dose CT scan to detect lung cancer as early as possible. But the insights from the project could also be useful for screening for other cancers, she said.

“We need to create structures to make sure that emerging and powerful tools like AI and machine learning are democratized,” she told Carmen. “Right now, if we just let the market decide who’s going to access these tools, they’re just going to benefit the rich, they’re just going to benefit urban areas and they're just going to benefit the white majority that doesn't have an accent,” she said.

What’s next: The text messages will be sent in the next month or so, and the project results will be ready in June.

 

Follow us on Twitter

Carmen Paun @carmenpaun

Daniel Payne @_daniel_payne

Ruth Reader @RuthReader

Erin Schumaker @erinlschumaker

 

Follow us

Follow us on Facebook Follow us on Twitter Follow us on Instagram Listen on Apple Podcast
 

To change your alert settings, please log in at https://login.politico.com/?redirect=https%3A%2F%2Fwww.politico.com/settings

This email was sent to edwardlorilla1986.paxforex@blogger.com by: POLITICO, LLC 1000 Wilson Blvd. Arlington, VA, 22209, USA

Unsubscribe | Privacy Policy | Terms of Service

No comments:

Post a Comment

The Best of Retail: Fashion Adapts to a Tough Reality

BoF is away for its annual break until 2nd January. In the meantime, we invite you to explore highlights of our news and analysis from t...