You'd probably start by using the LLM as a filter, not for the final diagnosis. If the LLM thinks something is strange (say, 5% of the time) --> refer to human

You'd probably start by using the LLM as a filter, not for the final diagnosis. If the LLM thinks something is strange (say, 5% of the time) --> refer to human doctor to ensure it isn't a false positive. If the problem is a shortage of doctors, this is one solution.