Eg. Combine the most promising 50 drugs, each at super low doses.
Low enough that side effects probably won't be bad, and if some prove entirely ineffective it doesn't matter.
But then, deliberately don't mix the ingredients precisely. Some people will get 10% more or less of each ingredient at random. Make precise records of this.
Now you have a natural experiment. Gradually increase dosage of elements that prove effective, and decrease ingredients which prove counterproductive - and you'll have a huge dataset with tens of millions of people wanting dementia treatment.
And probably a lot of death or maiming from drug interactions.
Which would also make it harder to get the couple million people to start with.
Edit: According to Claude you're actually talking more like several hundred million patients, i.e. more Alzheimer's patients than exist, to detect a single ingredient's effect.
With regard to the food supply, actually if there's anything we are collecting, it's the results (health outcomes). The bigger problem is the same problem as your proposed experiment: we aren't controlling the inputs, and so the results are entirely noise.