Writing in the medical journal The Lancet Digital Health, a team of MIT and Harvard researchers said an AI program could tell a patient’s ethnicity from X-ray and CT scan results with 90 percent accuracy.
That’s not good news, because even scientists haven’t figured out how these AI programs are able to distinguish races.
“When my graduate students showed me some of the results in this paper, I thought it must have gone wrong,” MIT assistant professor Marzyeh Ghassemi, one of the paper’s authors who analyzed the corresponding topic, told the media. When they told me, I really thought my students were crazy.”
Marzyeh Ghassemi
The article mentioned that the AI diagnostic system appears to diagnose and treat patients based on ethnicity, rather than the patient’s individual physical condition.
This practice will harm the patient’s health.
The researchers cited a case where an AI program had a higher chance of missing physical lesions in black and female patients when examining chest X-rays.
The purpose of this research is to confirm the extent to which AI systems can detect human race from medical images, and how they can detect race information from it.
To do this, the research team trained the AI system using medical images of different parts of the human body. The medical images fed to the AI system did not contain obvious ethnic markers such as hair texture, skin color, and BMI or bone density.
Through testing, the researchers found that the AI system was 90 percent accurate in identifying human races. The AI system can identify ethnic information from medical images of any body part.
Even more amazingly, the AI system was able to accurately identify ethnicity from even severely missing or damaged medical images.
What researchers are more concerned with is not the fact that AI systems can detect human race per se, but that the clinical performance of AI systems will be affected by these racial biases. And doctors may ignore errors in the diagnostic results of the AI system.
“The ability of AI to predict racial identity is not inherently important, but this ability is likely to be present in many medical image analysis models, which will exacerbate the problem of racial disparities already in the clinic,” the authors said.
Humans are currently unable to confirm which features of medical images an AI system detects a patient’s ethnicity. In addition, AI can identify a patient’s ethnicity from medical images of any part of the body, as well as from severely damaged medical images. This means that the use of medical imaging technology to create An AI system without racial bias would be very difficult.
Ghassemi told the outlet she guessed that perhaps the medical images recorded the levels of melanin in the patient’s skin in some unknown way, which was then identified by the AI system.
Based on the findings, it’s also possible that there are some innate differences between races.
Alan Goodman, a professor of biological anthropology at Hampshire College and one of the authors of “Racism Isn’t Race,” told the outlet he doesn’t quite agree with that statement.
Alan Goodman
In previous studies, scientists have struggled to find consistent racial differences in the human genome, but they have often been able to find consistent genetic differences based on the evolution of human ancestors. Therefore, genetic differences between people are more likely to be due to different characteristics of the evolution of individual human ancestors, rather than race.
Ghassemi said more research is needed on the issue to draw firm conclusions.
“We need to put AI systems on hold,” said MIT scientist and physician Leo Anthony Celi. “We can’t rush into hospitals and clinics until we can confirm that AI systems are not making racist or sexist decisions. .” Leifeng.com
Source: https://ift.tt/JUXd2kK
This article is reprinted from: https://www.leiphone.com/category/healthai/INDgICM30VlID6xF.html
This site is for inclusion only, and the copyright belongs to the original author.