It is very hard to build software like this which is both accurate and easy to use. Especially for general use. There are a lot of expert systems for medical diagnosis, including some which were accessible on the web (hundreds were at Revolution Health), but they are really not useful: they take far too long to use and are not sufficiently accurate. In fact Revolution Health removed most of them. One of the main problems is the terrible bandwidth of answering questions posed by a wizard.
But, at least, if it's on the iPhone, the huge costs of building and dealing with hardware are dealt with. I wish them luck, though I expect they are ignorant of the challenges ahead.
Some time ago, my strategy was to focus on high priority automated diagnosis, using much higher bandwidth channels, like microphones and cameras and so on. I had some machine learning algorithms for detecting skin cancer that I thought would significantly improve on the best existing algorithms, reaching 80-90% accuracy on a single viewing (higher than physicians which get around 60%), and nearing 100% if the moles were tracked. There were two major, but surmountable problems: I needed to do a lot of preprocessing to remove the variability of different cameras and shots and conditions, and I needed to get a good dataset to train and test (those are hard to acquire.) Ultimately I shelved the project because my cofounder wanted to work on something else (which I shelved because now I'm working on compressed air powered vehicles.)
How difficult were the regulatory issues around your software? I've been lead to believe that the FDA is quite difficult to work with on this type of software.
Furthermore, wouldn't legal liability be quite high?
The regulatory issues were very difficult, and I wasn't able to figure them all out before suspending the project. There wasn't an obvious precedent. My suspicion would be that if you built it, you could find some way to make money with it, but it would take years before the world would open up enough for people to, say, use it online.
It is very hard to build software like this which is both accurate and easy to use.
I would expect that a doctor that is used to dealing with the diagnosis process directly would be the best candidate to design one of this things. Although I'm sure there are quite a few challenges ahead.
though I expect they are ignorant of the challenges ahead.
He has a BS in Engineering and CS from a pretty good University, in addition to the MD. He should have at least some idea about what he is talking about, but this might just be wishful thinking on my part.
"There is no product like this available. The technology isn't difficult, just a bunch of algorithms organized properly: cardiovascular, neurological, orthopedics, etc."
Actually, people with good qualifications and lots of funding have been working on those algorithms for more than two decades. Products have been on the market for more than two decades (e.g. http://www.lcs.mgh.harvard.edu/projects/dxplain.html, http://scienceline.org/2008/01/04/doctor’s-diagnosis-version...). Almost all of these are not ready for use. Additionally medical experience is usually quite broad: not so focused on a particular disease. Some of the biggest successes in diagnostic process come from people specializing in diagnostic processes doing computer analysis (For example, Brenden Reilly's work on heart attack diagnosis, as popularized by Malcolm Gladwell's book, Blink).
But, at least, if it's on the iPhone, the huge costs of building and dealing with hardware are dealt with. I wish them luck, though I expect they are ignorant of the challenges ahead.
Some time ago, my strategy was to focus on high priority automated diagnosis, using much higher bandwidth channels, like microphones and cameras and so on. I had some machine learning algorithms for detecting skin cancer that I thought would significantly improve on the best existing algorithms, reaching 80-90% accuracy on a single viewing (higher than physicians which get around 60%), and nearing 100% if the moles were tracked. There were two major, but surmountable problems: I needed to do a lot of preprocessing to remove the variability of different cameras and shots and conditions, and I needed to get a good dataset to train and test (those are hard to acquire.) Ultimately I shelved the project because my cofounder wanted to work on something else (which I shelved because now I'm working on compressed air powered vehicles.)