3 ms·
I've been thinking about this for a while, in more general terms. Last year I had an issue with my eye. I bumped around from doctor to doctor (including one e
by colomon 14y ago
I've been thinking about this for a while, in more general terms. Last year I had an issue with my eye. I bumped around from doctor to doctor (including one eye specialist) for nine months, for at least half of that period on two different antibiotics and a ten minute a day eye cleaning ritual. Finally when nothing seemed to be working I was bumped up to an eye doctor at a bigger hospital two hours away. Within a minute of seeing me, he had a diagnosis. It involved no medicine at all, cleared up the problem within 48 hours, and still seems to be working fine nearly ten months later.
My thinking is that a properly programed medical diagnosis expert system probably could have figured out the issue by the second or third trip to the doctor. Further, my guess is that a nurse who has been properly trained to use such a system would be at least as accurate as your average doctor.
Unfortunately, my next guess is that any such system would be opposed by the medical establishment and quickly sued into non-existence by patients it made mistakes on (even if the rate of such mistakes was significantly below that of normal doctors).
- rksprst 14y agoYou're right physicians as a rule generally dislike software like that (partly because of their egos and distrust of software and partly because of job security). There are so many edge cases and so many examples physicians could give you of the AI/computer being wrong. A simple example is that any software will need input regarding the patient's history, his/her vitals. Right now heart rate data in ICUs is fed every second to storage systems. Let's say the this is a NICU and the baby's heart rate is suddenly elevated to extreme levels. The software would assume that something is wrong with the baby, when in reality a nurse is simply taking a blood sample and the baby started crying, being agitated by the nurse/needle. There are a lot of examples like this, it's a non-trivial problem to build use machine learning to take all of these factors into account to produce a diagnosis/treatment/intervention plan. I was recently at a medical/technology conference and the physicians in attendance were all actually in support of moving to using machine learning/AI (selection bias though, that was what the conference was about). They would rather have a machine with a 5% error rate than a human with 20% error rate. They believe that the role of physicians is going to change and become more of a management role as technology catches up.
- colomon 14y agoSounds like the mother of all selection bias, but still this makes me hopeful!
- spc476 14y agoSuch a system was developed in the 70s, MYCIN (http://en.wikipedia.org/wiki/Mycin http://en.wikipedia.org/wiki/Mycin). It wasn't used because of the state of technology at the time. Today it would probably perform much better.
- gamble 14y agoBack in the 90s, I was a research assistant to a couple of doctors who were working on that kind expert system. The main difficulty is the lack of electronic patient records. Most things in medicine are still done on paper. The other problem is that an expert system that does patient diagnoses is regulated by the FDA as a medical device, and it's incredibly involved to prove that it works well enough to be safe. That's why you see these systems pop up in research settings all the time, but very rarely deployed in the real world.