This type of application makes me a bit queasy, but given massive issues with access to healthcare I think this sort of use of LLMs is an inevitability. Is ther

This type of application makes me a bit queasy, but given massive issues with access to healthcare I think this sort of use of LLMs is an inevitability. Is there anything that can be done to make it more robust? Beyond the obvious issues with LLMs, confirmation bias seems like an especially big problem for a tool like this. It is seeing only the tests doctors ordered, only the notes doctors bothered to take, only the symptoms the patient thought were relevant, and so on. Would it come to the same conclusions if plausible synthetic results for other tests were included? Could it reliably predict results of a removed test given its previous conclusion? Could it predict the next action (test, referral, prescription) given results of a test? Is anyone working on this kind of adversarial approach to LLM outputs? Obviously LLMs are far from ideal for an application like this but could this type of probing for internal consistency improve the error rate and allow for some measure of confidence?