I don’t see this argument’s strength. Either the MR is indicated in a particular workup and the physician didn’t realize this or the MR isn’t indicated and won’t be performed (the physician will review the AI recommendation, the radiologist will refuse the study, or the LLM used by the payout will reject the prior auth).
The argument’s natural extent is that we should not do any more testing than we currently do if we begin using LLMs, which has thr problem I outline above.
APP/NP/PA’s often order more imaging (there are studies to this effect) but this is potentially due to their education not focusing on when tests should not be pursued as opposed to carrying out what they’re asked to do by physicians. This is the setting where Opus/Claude may drive higher utilization (subject to the gates I highlighted earlier).
gotta be careful about assigning cause and effect like this; I don't just mean 'don't anthropomorphize the model,' but that one might draw false conclusions on 'why' something happens in a system
The underlying issue is that these models tend to shotgun orders because, at their core, they want more data. More data isn't always better for the patient: it generates incidentalomas, follow-up imaging, and downstream workups etc that carry their own risks and costs.
I do like the agent triaging in basket messages, though. That's pretty clever. Where it breaks down is when things get super specialized. This works well for primary care, etc. Maybe not in your ultra/hyper specialized setting.