I'm not sure I see that as a fatal flaw... ultimately, the user is sitting in front of a digital box that can guide and prompt them in all the same ways a researcher can--the only limit is in performing tests that require a medical professional to assess biomarkers. If the pace of medical device development continues, it's even reasonable to think something like a Theranos-that-works could commoditize the process while being intrinsically tamper-resistant.

Regardless, users can be prompted to perform any software action (knowingly or unknowingly, to affect bias) and that action can be measured by the system. It may so happen that every critical measurement occurs unbeknownst to the user, before they self-report anything (if at all). As we are currently undergoing a period of sensor-proliferation (fitness/health devices, wearables, internet of things, etc...) it's not unrealistic to think we will soon be able to instantly correlate data from a smartphone camera, blood/tissue, and the cloud.

Now there's always the problem of intentional fraud/deception, but I think the aggregate nature solves that problem. A small percentage will try to "break" the system, and that small percentage will never surpass a critical threshold with enough volume. In terms of ML/SVM's, we're now very good about filtering outliers or "misrepresented data"... while the responsibility is on you to develop a reliable classifier (for data-consistency more than arbitrary measurement), I imagine at scale you could infer trends with the same relative accuracy of traditional academia and research.

It's a really fascinating new direction--even if only an adjunct to traditional research--and I'll definitely be keeping an eye on the project.