Wrote this last yr. before jumping back into clinical ML but never got around to sharing it. Added some 2019 updates (clinical BERT and approach for industry applications).
I updated the appendix with a few papers from 2019 but it felt like there were 10X more papers compared to 2018 (which is fantastic!). But, instead of making the appendix even longer, I highly recommend just following http://www.arxiv-sanity.com/search?q=health to stay up-to-date.
Even if you’re not in the health space, you’ll find ingenious interpretability techniques and tips leveraged by researchers out of necessity of being in the clinical space. I conclude with a realist note on the challenges that lie ahead for safely transitioning research to the clinical setting.
Great work! I've got a rather wild ML ERH story to share.
So wrote my thesis on the design of EHR system (http://barnett.surge.sh/), and I interviewed this guy who said he was working on a crazy revolutionary ML-powered EHR system. It apparently had features such as voice recognition for writing notes instead of typing - but it also featured taking unstructured medical notes, and with the great AI, structuring them. So suddenly large scale medical record analysis would be easy - a rather incredible idea. However it sounded like they were at a very early stage, and they had nothing to show - so I thought 'yeah good luck getting that working' and trucked along with the thesis.
Turns out a journalist had heard of this medical AI, and dug way deeper. This guy (Albi) had convinced the GP I talked to that he had a working medical AI system. He had actually gone so far as to have the GP email 'the ai' and have it reply. However the whole thing was a fucking sham - it didn't exist, and the person replying was honestly probably just Albi. They were trying to raise funding to get it further, based on this claim of a 'functional ai'.
I don't think the GP I interviewed was in on the con, I think he was being taken for a ride. However once the article was published, there was heaps of attention and it all kind of fizzled out, and Albi was recently bankrupted.
But, I'm glad to see somebody is working on this problem - who isn't a conman haha.
Every major EHR vendor and some other related organizations have been working on that problem for years. It's seen as a "holy grail" of EHR functionality.
I updated the appendix with a few papers from 2019 but it felt like there were 10X more papers compared to 2018 (which is fantastic!). But, instead of making the appendix even longer, I highly recommend just following http://www.arxiv-sanity.com/search?q=health to stay up-to-date.
Even if you’re not in the health space, you’ll find ingenious interpretability techniques and tips leveraged by researchers out of necessity of being in the clinical space. I conclude with a realist note on the challenges that lie ahead for safely transitioning research to the clinical setting.