Addressing cognitive bias in medical language models