I'm kind of apprehensive about this advancement. There's many ways to mess this up, from intentional dataset poisoning (to encourage patients to get unnecessary procedure) to just terrible errors made with just a shift in data capture methodology. This paper has a lot of detail on how such systems can fail (https://arxiv.org/pdf/1804.05296.pdf).
Unfortunately, it's a multi-trillion dollar industry with insanely entrenched and deep pocketed players, and the old boy network the likes of which us software people can't even imagine. And that's before you consider the regulatory framework and the cost of compliance with it.
Solving medical errors would be huge. The problem is that it's a long-tailed problem with no clear way to solve with software. Every error is different, and there are already lots of bureaucratic mitigations in place every step of the way. There are no algorithms capable of the general reasoning required to broadly spot all types of errors including de novo ones.
The only way to truly guarantee no medical errors is to replace every agent in a hospital with a machine.
I bet just even solving the errors in diagnostics and/or catching stuff before it's too far gone would be huge in itself. I'm not talking AGI here, just bog standard perceptual stuff: looking at xrays, MRIs, mining medical records for patterns, handing unstructured medical records better, low hanging stuff like that.
Automating literally any automatable healthcare process with software is virtually guaranteed to eliminate errors in that process, because it wouldn't be implemented in the first place if its accuracy wasn't superhuman. It's a lot of different bog standard stuff like you said. Medical imaging, yeah this should already have been done by now. Parsing medical records--you don't even need neural networks for that. But then it's no longer a Solving Medical Errors problem but instead a Automating Medical Job X problem. The general problem per se as I envision it would watch a stream of written actions performed in a hospital (maybe even surveillance footage) and determine if an error happened or is likely to happen; that's AGI.
I'm kind of apprehensive about this advancement. There's many ways to mess this up, from intentional dataset poisoning (to encourage patients to get unnecessary procedure) to just terrible errors made with just a shift in data capture methodology. This paper has a lot of detail on how such systems can fail (https://arxiv.org/pdf/1804.05296.pdf).