Over the past year, healthcare experts have identified ways in which artificial intelligence, and similar tools, reinforce existing biases involving race or ethnicity.
Now, many are taking concrete steps to eliminate these biased AI models.
A team of doctors at the American Academy of Pediatrics is going through hundreds of algorithms and other materials to find, and remove, anything that could exacerbate racial disparities, a “herculean task,” as the JAMA report describes it.
The potential for AI-enabled bias could be a major problem for diagnosing life-threatening conditions in seniors. And within senior living communities, there is actually a slightly higher percentage of African Americans compared to the general US population, one study notes, though the report adds that this is true for nursing homes and not assisted living facilities.
The AAAP review is one of several such attempts to vet clinical algorithms, the story notes.
The Doris Duke Foundation gave the AAAP and several other organizations $10 million in grant funding to identify such racial biases.
The Council of Medical Specialties Societies also recently released a report on using algorithms and how to improve them to eliminate racial bias.
These tools can reflect bias in two different ways, the US Office of Minority Health explains: One, AI can be trained on pre-existing data sets in which minorities are often underrepresented. Secondly, algorithm developers can directly embed the tools with implicit assumptions, such as symptom occurrence, that cause problems.
While, in theory, healthcare providers relying on AI for clinical diagnoses should be reviewing and double-checking results, the reality is that most clinicians will simply double down on any inaccurate AI diagnoses, recent research shows.
One example of a course correction hospitals had to make was regarding kidney function estimates, according to a 2020 JAMA report. Certain tools were contributing to delayed kidney transplants for African American patients with kidney disease, the report found.
