Artificial intelligence software is being developed that can read x-rays, and other medical scans. This software has the potential to spot lung cancers that doctors often miss. According to a new study, these algorithms can also detect something that doctors don't look for in scans like x-rays and other medical scans.According to the study's authors, and other AI experts in medical medicine, it is more important than ever that health algorithms are able to accurately predict race. This task is made more difficult by the fact that the authors aren't certain of what cues they used to predict a person's race.Tests on five types images used in radiology research (e.g., chest and hand xrays and mammograms) have shown that algorithms can detect race. Images were taken from patients who identified themselves as Black, White, and Asian. The researchers used images that were labeled with the patient's self-reported race to train algorithms for each type of scan. They then challenged the algorithms to predict patients' race in unlabeled images.Radiologists don't usually consider a person's racial identity, which isn't a biological category, to be visible on scans that look below the skin. The algorithms were able to detect it in all three racial categories and from different angles of the body.The algorithms were able to identify the identity of a Black person in most cases of scans more than 90% of the times. The worst algorithm was able to identify the Black person in the scans 80 percent of times. The best algorithm could be identified as being 99 percent correct. Although the results and associated code were uploaded online by more than 20 researchers who are experts in medicine and machine-learning, peer review has yet to be done.These results raise concerns about AI software increasing inequality in healthcare. Studies show that Black patients receive less care than whites and those who are wealthy.Machine-learning algorithms can be trained to read medical images by being fed many examples of diseases, such as cancer. The algorithms can identify patterns in pixels statistically associated to these labels by digesting many examples. For example, they can determine the shape or texture of a lung nodule. These algorithms can detect skin cancers and other diseases that are invisible to humans, which is why they have been compared to doctors.Judy Gichoya is a Emory University assistant professor and radiologist who participated in the new study. She says that the revelation that image algorithm can detect race in internal scans may make it easier for them to learn other inappropriate associations.We need to educate the public about this issue and find ways to reduce it. Judy Gichoya is a radiologist at Emory University and an assistant professor.Due to socioeconomic and historical factors, medical data used to train algorithms can often contain traces of racial disparities in diseases and treatment. An algorithm could use scan patterns to determine a patient's race and then suggest diagnoses that are racially biased. This would not only be the obvious medical anomalies that radiologists search for, but also patterns that may correlate with those patterns. This system could give false positives or an incorrect diagnosis to some patients. A different algorithm could suggest different diagnoses to a Black person or a white person with similar symptoms.Gichoya states that we need to educate people about the problem and find ways to reduce it. Her collaborators included representatives from Purdue and MIT, Beth Israel Deaconess Medical Center (MIT), Beth Israel Deaconess Medical Center (National Tsing Hua University, Taiwan), University of Toronto, Stanford, and Stanford.Studies in the past have shown that medical algorithms can lead to biases when it comes to care delivery. Image algorithms may also be less effective for certain demographic groups. A widely-used algorithm that prioritizes care for the most sick patients was discovered to be disadvantageous for Black people in 2019. Researchers at MIT and the University of Toronto found that algorithms used to detect conditions like pneumonia on chest xrays can sometimes perform differently for different sexes and ages.
