[ad_1]
Millions of dollars they are spending to develop Artificial intelligence software that reads x-rays and other medical tests that can detect things that doctors are looking for but sometimes lose, such as lung cancer. A new study has reported these algorithms he can also see something that doctors don’t look for in such sweeps: the patient’s race.
The study’s authors and other AI medical experts say the results make it more crucial than ever to verify that health algorithms work well for people with different racial identities. Complicating this task: the authors themselves do not know what axis the algorithms created to predict a person’s race use.
Evidence that the algorithm can read the race through medical examinations of people came from tests of five types of images used in radiology research, including chest and hand radiographs and mammograms. The images included patients who identified blacks, whites, and Asians. For each type of scan, the researchers trained algorithms using images labeled with the patient’s own race. They then challenged algorithms to predict the race of patients in different unlabeled images.
Radiologists do not generally see a person’s racial identity — which is not a biological category — visible in scans that look under the skin. However, it was somehow demonstrated that algorithms are capable of accurately detecting three racial groups and different views of the body.
In most types of scans, the algorithms were able to correctly identify which of the two images was more than 90 percent of a person’s time in black. The worst-performing algorithm has also been successful 80% of the time; the best was 99 percent correct. The results and associated code They were published on the Internet at the end of last month by a team of more than 20 researchers specializing in medicine machine learning, but the research has not yet been reviewed.
The results have raised new concerns that AI software could exacerbate inequalities in health care. Studies show that black patients and other excluded racial groups receive less attention than wealthy or white people.
Automatic learning algorithms are symptomatic for reading medical images, feeding many labeled examples of conditions such as tumors. By digesting many examples, algorithms can study pixel patterns that are statistically associated with tags, such as the texture or shape of a lung nodule. Some algorithms did so by rival drugs to detect cancer or skin problems; there is evidence that they can detect signs of disease invisible to human experts.
Radiologist Judy Gichoya, an assistant professor at Emory University who has worked on the new research, says the imaging algorithms reveal the possibility that they can “see” the race in internal scans to learn about inappropriate associations.
The medical data used to train the algorithms show traces of racial differences in diseases and medical treatments due to historical and socioeconomic factors. This could lead to an algorithm for scanning statistical models to look for statistical patterns in the patient’s race, suggesting diagnoses related to race-biased models from his training data, not just the spectacular medical abnormalities sought by radiologists. This system can give some patients a misdiagnosis or a false light. An algorithm may suggest different diagnoses for blacks and whites with similar signs of disease.
“We need to educate people about this issue and we need to research what we can do to alleviate that,” Gichoya says. Project collaborators came from organizations such as Purdue, MIT, Beth Israel Deaconess Medical Center, Tsing Hua National University in Taiwan, the University of Toronto, and Stanford.
Previous research has shown that medical algorithms have led to biases in care and that image algorithms can lead to imbalances in different demographic groups. In 2019, a widely used algorithm for prioritizing care for sick patients was discovered disadvantages Blacks. In 2020, researchers at the University of Toronto and MIT showed that algorithms trained to mark conditions such as pneumonia on chest radiographs were different for people of different sex, age, race, and health insurance types.
[ad_2]
Source link
