On the one hand, using voice as a pre-screening test in places where the normal screening test is too expensive to administer routinely seems like a great thing. i.e.: Read this paragraph to the machine, and we’ll figure out whether it’s worth actually testing you for T2DM, Parkinson’s, stomach cancer, lung cancer, etc, etc. If that substantially reduces the number of tests administered without making too many false negatives, then you can really improve health in some very poor areas.
This data set is definitely not going to give that. It’s not even particularly compelling evidence that it’s possible. It is, IMO, compelling enough to study further. Bigger sample sizes, fewer than 84 recordings over 2 weeks. It kind of looks like p-value chasing, and running a bigger study would answer that.