News for hackers. Stuff that matters.
Re:Requests for Startups Refreshed by Cromatico on Tuesday September 27, 2016 @07:02PM (#12594007)
An idea for mixing AI with healthcare, using Tensor Flow with a huge image database of skin conditions so the user takes a pic on a phone and it determines the illness and the cure.
Re:Requests for Startups Refreshed by TuringNYC on Tuesday September 27, 2016 @09:32PM (#12594778)
The technical aspects can be de-risked fairly quickly here -- check out the ISBI dataset http://biomedicalimaging.org/2016/?page_id=422 and all the papers around it.

The societal challenges will be a devil to defeat. FYI I actually quit my Wall St job several years ago, did something similar with diagnostic imaging, built a working classifier that is doctor-level sensitive (low false negative especially). For us it was near impossible to succeed in the US because the US healthcare system is designed from all directions to be inefficient.

Our tepid success came from moving overseas temporarily and proving out the system in a smaller country with a forward-looking healthcare system with less conflicts of interest. We focused on a local problem (Tuberculosis.) Even then, I'm almost two years in and ended up acquiescing to a premature acquisition because the healthcare system is so difficult to navigate. My lockup is almost complete, the next startup, I'm going to focus on something easier like world peace.

I don't mean to discourage you, but you would likely be a lot more successful if you have some "in" with a hospital system willing to work with you. IMHO, in healthcare, the technology is the easiest part.

Incidentally...yes...we've built it on TensorFlow! And if you do wish to talk tech about this idea, feel free to reach out to me.

Re:Requests for Startups Refreshed by spitfire on Tuesday September 27, 2016 @11:40PM (#12595285)
I'm doing this right now actually. The risk with using ML - in particular NN methods, in health care is that you have to be able to explain and support your conclusions.

Also depending on the training set used you can get some VERY bad results.

https://vimeo.com/125940125

Re:Requests for Startups Refreshed by sbuccini on Tuesday September 27, 2016 @07:18PM (#12594121)
I've looked into this. There are actually a couple of attempts that use basic heuristics for detecting melanoma such as color, symmetry, and size. However, it's harder here in the US since you have to get approved by the FDA in order to make a "diagnosis".