Re:Ask YC: This is our tech, how can we make this into a pro...
by anmol on Sunday October 17, 2010 @04:07PM
(#1801269)
The 90% is a misquote, good catch.
The actual accuracy is between 60-90%, depending on clusters of symptoms, and apparently that was stripped somewhere during editing. Oh well.
Re:Ask YC: This is our tech, how can we make this into a pro...
by andrewljohnson on Sunday October 17, 2010 @04:17PM
(#1801294)
With a range that large, I would be highly suspect that when you did get 90%, it was just fitting the data. And how would you even confirm if they had the flu - even a doctor could misdiagnose mild flu, which adds more noise.
60% really means very little if it's a binary choice - flu or no. You'll be right 50% of the times no mater what.
Re:Ask YC: This is our tech, how can we make this into a pro...
by moultano on Sunday October 17, 2010 @04:20PM
(#1801308)
>60% really means very little if it's a binary choice - flu or no. You'll be right 50% of the times no mater what.
What? Just because something is binary doesn't mean it happens half the time.
Re:Ask YC: This is our tech, how can we make this into a pro...
by andrewljohnson on Sunday October 17, 2010 @04:22PM
(#1801316)
Hmm, yeah... good point on that. I'll leave my comment, but I take it back :)
Re:Ask YC: This is our tech, how can we make this into a pro...
by anmol on Sunday October 17, 2010 @04:36PM
(#1801348)
With severely unbalanced classes, overall accuracy is not a good metric. So we use recall of the symptom class, and how it varies with the precision. Recall = 0.6 to 0.9. All results are based on cross-validation.
There is a lot more sophistication we can add (e.g. markov properties), these are almost first-order results.
Nonetheless important to validate with other populations. That would be a big part of testing any product.
I'm very surprised by this - it strikes me that I'd expect it to be hard to tell from such data that it wasn't a stomach upset, tiredness, onset of disease or simply a heavy cold. Indeed 90% sounds like it might be higher than correct identification via an interview with a doctor.
>[...] in Rwanda between 2006 and 2009. He saw a clear reduction in people's movement, which may have been due to the disease. But the outbreak was caused by floods, which also limited mobility. Distinguishing between the two possible causes on the basis of phone data alone was impossible, he says.
Makes me the 90% figure highly suspect.
That aside I wonder if it wouldn't be easier to have a smart phone app that gives you epidemiology info in return for entering your health status for the day. To encourage continued use a sort of diary function and stats could be added. Wouldn't this be more predictive than monitoring movement, and less intrusive?