The hard cases, which would be useful to a doctor, occur very rarely. My father's unusual reaction to a post-bypass drug regimen was something like the 3rd time that happened in Canada. How do you "train" that into a neural network?
It's super difficult to beat something like linear regression in this sort of thing (ideally combined with domain knowledge -something neural approaches mostly fail at), and even linear regression gives awful results.
To make this concrete: I used to do a lot of HIV testing and counseling. Whenever I'd ask someone "how often do you use condoms?" they'd 100% of the time say "every time!" When I'd then switch to asking "when was the last time you didn't use a condom?" they'd often reply along the lines of "last week."
This kind of issue happens not just with awkward questions, but also with more "objective" data like labs and medications.
- Was the antibiotic script written for a patient necessary for their condition, or was the physician tired and couldn't fend off a particularly assertive patient who was convinced antibiotics would solve their viral infection?
- Are labs randomly ordered, or ordered in a targeted fashion based on the "hunch" of a physician? If the latter, then the presence of lab result would itself be correlated with having some disease, and could thus throw off your entire model.
- How does the patient's insurance status influence the tests they have ordered? At one clinic I'm aware of, they typically order HIV/GC/Chlam screenings as a bundle, but if the patient has a PPO insurance they are also more likely to throw in a syphilis screening.
- How does patient preference factor into what gets diagnosed, ordered, and treated? You might have two early stage, equal risk breast cancer patients, one whom opts for a double mastectomy because she saw what happened to a coworker who had breast cancer and doesn't want to risk recurrences, whereas the other opts for a lumpectomy because she thinks her risks are low enough to be more conservative in treatment.
These issues are all solvable, but for AI to have a useful impact in healthcare, there's a lot more work than just throwing a bunch of data into a deep learning model.
Except for some really common problems, like cases of influenza, every patient is unique.
AI can be used to handle simpler problems - identifying patients at risk of repeat admissions, or flagging the likelihood of dehydration for example.