How gender biases can trip up AI diagnostics

New research highlights what appears to be a pervasive challenge in building AI models to diagnose disease: gender disparities. The study found that when women were underrepresented in or excluded from the patient pool on which the machine-learning model was trained, the subsequent algorithm performed worse in diagnosing them with a range of medical conditions. The study, researchers say, demonstrates how biases can sneak into computer models, and shines a light on an issue that has broad implications. Researchers previously reported that a predictive model for kidney function decline performed worse for women, who only made up 6% of patients whose data trained the algorithm. More here.