(OP here) I worked on one of the first studies on detecting diabetes from wearables in 2018. It was kind of fun to review how the field has evolved. The basic approach is the same (pretrain a self-supervised foundation model on sensor data, fine-tune for a particular application) but the scale is several orders of magnitude larger.

It seems their algorithm wouldn't work for me -- it would rate me falsely insulin insensitive:

  > " Google’s study found the expected pattern: HOMA-IR rose with resting heart rate (r = 0.27) and fell with daily steps (r = -0.25) and HRV (r = -0.14)."

My actual HOMA-IR values were so low they were _below_ the reference range, the exact opposite of what this model would predict:

https://i.imgur.com/ISEwgAl.png