(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.