Saturday, May 18, 2024

From medical pictures, artificial intelligence determines a patient’s race

AI may detect self-reported race from medical photos that are not observable by human specialists, according to a study. Miseducation of algorithms is a severe issue; when artificial intelligence reflects the unconscious attitudes, bigotry, and biases of the humans who created these algorithms, serious harm can result. For example, computer programmes have incorrectly identified Black defendants as twice as likely to reoffend as white offenders. When an AI employed cost as a proxy for health needs, it incorrectly identified Black patients as being healthier than equally ill white patients since they received less money. Even artificial intelligence (AI) used to compose a play relied on detrimental preconceptions for casting.

There are numerous examples of bias in natural language processing, but MIT researchers have looked into another essential but understudied modality: medical imaging. The researchers discovered that AI can accurately estimate patients’ self-reported race from medical photos alone, using both private and public databases. The scientists trained a deep learning model to classify race as white, black, or Asian using imaging data from chest X-rays, limb X-rays, chest CT scans, and mammograms – even though the pictures themselves had no explicit indication of the patient’s race. Even the most experienced physicians can’t accomplish this, and it’s unclear how the model did it.

The work was published in Lancet Digital Health on May 11 under the title “AI recognition of patient race in medical imaging: a modelling study.” Celi and Ghassemi collaborated on the paper with 20 other authors from four different countries.

To begin the tests, the researchers demonstrated that the models could predict race across a variety of imaging modalities, datasets, and clinical activities, as well as across a variety of academic centres and patient groups in the United States. They employed three big chest X-ray datasets and evaluated the model on a previously unseen subset of the training dataset as well as a completely other dataset.

Differences in physical characteristics between different racial groups (body habitus, breast density), disease distribution (previous studies have shown that Black patients have a higher incidence of health issues like cardiac disease), location-specific or tissue-specific differences, effects of societal bias and environmental stress, and the ability of deep learning systems to detect race when mulling over data were all explored in an attempt to explain the model’s behaviour.