Study Suggests AI Models Could Be Relying on ‘Demographic Shortcuts’ During Medical Diagnosis

New research has determined that artificial intelligence (AI) models may be depending on “demographic shortcuts” when making medical diagnoses. Prior studies had determined that these models could reliably predict a patient’s race from their chest X-rays, something even the best radiologists couldn’t do.

This has been further put under the spotlight by this recent study, which determined that these artificial intelligence models have discrepancies in their ability to reliably diagnose images of individuals of different races or genders.

For their study, the researchers used chest X-ray datasets to train the models to predict if patients either suffered from a collapsed lung, fluid buildup in the lungs or enlargement of the heart. While the models performed well generally, a significant number displayed discrepancies between accuracy rates for women and men, as well as for Black and white patients. This meant that these models’ diagnostic assessments produced incorrect results for Black people and women, among other groups.

The researchers also observed that the models were able to predict the race, gender and age of the X-ray subjects. Additionally, they noticed a correlation between the accuracy of each model in making predictions based on demographic data and the size of its fairness gap.

Senior author of the study, associate professor Marzyeh Ghassemi of electrical engineering and computer science at MIT, stated that it was well-known that machine-learning models were good forecasters of human demographics. He explained that the research showed the models’ capacity then related this capacity to underperformance across distinct groups.

Ghassemi is a member of the Institute for Medical Engineering and Science at MIT.

The researchers also discovered that the models could undergo training again to help enhance their fairness. The researchers’ approach to making the models less biased worked optimally when they were tested on the same patient types they were trained on. In their report, the researchers noted that they observed that when these models were used on other patients, the fairness gaps resurfaced.

One of the lead authors of the study, Haoran Zhang, stated that the primary takeaways from their study was that external models needed to be assessed on their own data as results may not transfer to another population. Additionally, Zhang added, models needed to be trained on developers’ data if adequate data was available.

Other authors of the paper include associate professor of radiology and imaging sciences Judy Gichoya, of Emory University School of Medicine; Dina Katabi, a professor of electrical engineering and computer science at MIT; and Yuzhe Yang, a graduate student at MIT.

Until these kinks in the AI models powering some of the newer forms of medical diagnostics systems are ironed out, the market may have to rely more on the tried and tested diagnostic tools from companies such as Astrotech Corp. (NASDAQ: ASTC) to establish the different ailments afflicting the patients who seek medical help for the symptoms they exhibit.

NOTE TO INVESTORS: The latest news and updates relating to Astrotech Corp. (NASDAQ: ASTC) are available in the company’s newsroom at https://ibn.fm/ASTC

About BioMedWire

BioMedWire (“BMW”) is a specialized communications platform with a focus on the latest developments in the Biotechnology (BioTech), Biomedical Sciences (BioMed) and Life Sciences sectors. It is one of 60+ brands within the Dynamic Brand Portfolio @ IBN that delivers: (1) access to a vast network of wire solutions via InvestorWire to efficiently and effectively reach a myriad of target markets, demographics and diverse industries; (2) article and editorial syndication to 5,000+ outlets; (3) enhanced press release enhancement to ensure maximum impact; (4) social media distribution via IBN to millions of social media followers; and (5) a full array of tailored corporate communications solutions. With broad reach and a seasoned team of contributing journalists and writers, BMW is uniquely positioned to best serve private and public companies that want to reach a wide audience of investors, influencers, consumers, journalists and the general public. By cutting through the overload of information in today’s market, BMW brings its clients unparalleled recognition and brand awareness.

BMW is where breaking news, insightful content and actionable information converge.

To receive SMS alerts from BioMedWire, “Biotech” to 888-902-4192 (U.S. Mobile Phones Only)

For more information, please visit https://www.BioMedWire.com

Please see full terms of use and disclaimers on the BioMedWire website applicable to all content provided by BMW, wherever published or re-published: https://www.BioMedWire.com/Disclaimer

BioMedWire
San Francisco, CA
www.BioMedWire.com
415.949.5050 Office
Editor@BioMedWire.com

BioMedWire is powered by IBN