U. expert discusses AI bias in research studies
In a recent study, researchers determined that the field of studying fairness and bias in healthcare-related artificial intelligence (AI) is largely dominated by white and male researchers.
The study concluded that white and male researchers' scholarly works are cited disproportionately more often than other races and sexes, mainly in first and last authorship.
The order of authorship indicates their contribution to the study, with the last author being the primary contributor to the research. The study also found that authors from wealthier countries were more often referenced in major journals than their peers.
Charles Senteio, an author of the study and an associate professor in the School of Information and Communication, said the study aimed to measure the diversity of authorship in health AI research.
The study compiled 1,614 articles related to AI fairness to analyze the diversity of authorship and filtered the articles for eligibility. Articles examined in this study discussed machine learning fairness, were related to healthcare and included the clinical applications of their findings.
For the 1,984 authors studied, the researchers created distributions of each author's ethnicity and gender, nationality, citations and funding received. Sixty-four percent of the total authors were white, 60 percent of the total authors were male and white male last authors accounted for 58.3 percent of the total citations, according to the study.
Additionally, only .5 percent of all authors came from low-income countries, according to the study.
Senteio said the study itself faced limitations and is yet to be peer-reviewed. Additionally, he said that measuring diversity in the context of research can be fraught with complications.
"Even collecting the race of patients today is not always straightforward because it's not done in a uniform way. For example, sometimes, it's already in the record. We don't know necessarily how it got there," he said. "So simply asking a patient their race and to indicate it. That's the simplest, the most common way, but there are many reasonable patients who would rather not disclose that."
Additionally, Senteio said that having diverse teams does not inherently guarantee that the research conducted will have no bias.
Despite these limitations, he said that researchers will not understand how bias negatively impacts health AI until they start measuring aspects like the research's authorship.
"I wouldn't say necessarily that mandating racial equity among authors for all work in AI bias regarding health care is a solution. But we don't even know what a problem is right now. So let's start somewhere and see what we get … and see what researchers think and see where it goes," Senteio said.
He said an issue recently found relating to health AI bias involves the pulse oximeter device. The technology, which measures the level of oxygen in the blood, was found to read dark skin differently than light skin.
Additionally, Senteio said that AI bias had been found in radiology, where X-rays traveling through darker skin look different than X-rays that do not, which can affect how radiologists examine these scans.
In terms of working with these medical devices, Senteio said that developers should factor in the potential differences between patients' skin color when designing this technology.
"How do we know that when we develop guidelines from these algorithms that they'll actually be appropriate for the patients who are going to be subjected to some of these recommendations? Maybe having diverse teams will put us in a better position to act," he said.
Regarding the Rutgers community, he said that he hopes he and his co-authors' research will inspire people to think of diversity as a method of producing better research, not just as a quota to hit.
"Diverse teams, which doesn't just mean representation, but empowerment and safe spaces to actually speak up and put forth points of view that might not be aligned with the mainstream. That might not be aligned with what leaders or (principal investigators) of projects are putting forward," Senteio said. "I think that that kind of thinking and realization would go a long way to some of the very noble DEI efforts that we see almost everywhere."