alt Hacker NewsApart from the likely technical infeasibility of your idea in today's society, this would require a humongous and diversified population sample to be meaningful (your 'heterogeneous bits'). This follows directly from the complexity of metabolic pathways you wish to analyze. Socially, you'll only be able to achieve that by not asking your sample for consent. Otherwise you'll have a highly biased sample, which could still be useful but for severely restricted research questions.
There are some pretty big longitudinal studies with consent ( "45 and up" are a quarter of a million people, for example - that's big enough that working predictions within the cohort would be a worthwhile health outcome).
There are nevertheless privacy issues, which I did not address as my first comment was already very long, especially for a tangent. Most obviously, people would be consenting to the collection of data whose significance they cannot reasonably forsee.
I do agree that most current AI companies are unlikely to be a good steward of such data, and the current rush to give away health records needs to stop. In a way it's a good thing that health records are currently so limited, since the costs will so obviously outweigh the benefits.