If race can be guessed based on skeleton, then “bias” could also be interpreted as “giving useful medical signal based on measurable traits”. In order to eliminate human bias, wouldn’t you’d need to strip the “race” info from the notes given to the radiologist?<p>It is not clear at all whether what the ML is picking up is from genetics or from other sources (e.g. diet and other practices that associate with the cultures that often follow along with visible markers of race). Getting the ML model to explain itself will be an interesting follow-up to watch.