In defence of post-hoc explanations in medical AI