AI can diagnose type 2 diabetes in 10 seconds from your voice
diabetes.co.uk
diabetes.co.uk
Fat affects vocal chords, and this can be detected in the vocal patterns.
This whole "AI Tool" could be replaced with a simple question:
"Are you fat?"
But people often aren't honest with themselves about this topic.
Roughly 72 per cent of the participants (79 women and 113 men) had already been diagnosed as nondiabetic. The other participants (18 women and 57 men) had been diagnosed with type 2 diabetes.
All participants recorded a phrase six times per day for two weeks, resulting in a total of 18,000 recordings. The scientists then pinpointed 14 acoustic differences between those with and without type 2 diabetes."
For a clinician, sensitivity and specificity are much more useful. It's too bad they didn't publish these.
That is to say, don't compare to the accuracy of an even coin flip; compare to the accuracy of a coin flip that flips according to the proportion of the people with the attribute you care about.
I don't think that's right. Always answering yes is going to have 75% accuracy if the prevalence is 75%, but always answering no will have 25% accuracy, and the coin flip will pick both of those equally (and uncorrelated with whether the individual has diabetes or not).
"Arrest the witch!!"
AI to Diagnose Diabetes Based on Seconds-long Snippet of Voice:
"Uh. Interesting ...!"
Note that the “optimal model” includes age and BMI.
The predictive model seems to be logistic regression and Naive Bayes. There’s no fancy AI here. Just some basic feature summary.
I don’t have time to go into it too detailed but I’m noticing they’re saying some vocal features are P <0.001 for the matched datasets even though the values across the two classes seem awfully similar. I’m not sure what’s going on there.
I _quickly_ checked the study [1] and didn't see anything obvious.
Given that our voices reveal so much about the state of a human (think of mood, emotions), I think it's at least possible that voices could reveal something about metabolism. Just we cannot hear it because hearing diabetes is not as useful as hearing something about the mood of another person.
[1] https://www.mcpdigitalhealth.org/article/S2949-7612(23)00073...
It's just because I don't think this is evidence enough to change my priors. i.e. My gut tells me this is wrong and I don't buy it.
The sample is 267 people, tiny enough that I'd expect an analysis to be done with a linear model with few features. They used 14 features, but had a pipeline to select "model (out of 3), feature set, and threshold for prediction".
There's _so_ many degrees of freedom there that _by default_ I assume there's leakage.
I'd love to see this paper replicated! It would be amazing if this were true. But if I had to bet on this, I wouldn't give more than 20% chance of being true.