Trendy wearable gadgets can conveniently file varied biosignals within the many various environments of each day residing, enabling a wealthy view of particular person well being. Nonetheless, not all biosignals are the identical: high-fidelity biosignals, equivalent to photoplethysmogram (PPG), include extra physiological info, however require optical sensors with a excessive energy footprint. Alternatively, a lower-fidelity biosignal equivalent to accelerometry has a considerably smaller energy footprint and is on the market in virtually any wearable gadget. Whereas accelerometry is extensively used for exercise recognition and health, it’s much less explored for well being biomarkers and prognosis. Right here, we present that an accelerometry basis mannequin can predict all kinds of well being targets. To realize improved efficiency, we distill representational data from PPG encoders to accelerometery encoders utilizing 20 million minutes of unlabeled knowledge, collected from ~172K members within the Apple Coronary heart and Motion Examine underneath knowledgeable consent. We observe robust cross-modal alignment on unseen knowledge, e.g., 99.2% top-1 accuracy for retrieving PPG embeddings from accelerometry embeddings. We present that distilled accelerometry encoders have considerably extra informative representations in comparison with self-supervised or supervised encoders skilled straight on accelerometry knowledge, noticed by no less than 23%-49% improved efficiency for predicting coronary heart fee and coronary heart fee variability. We additionally present that distilled accelerometry encoders are readily predictive of a wide selection of downstream well being targets, i.e., they’re generalist basis fashions. We imagine accelerometry basis fashions for well being might unlock new alternatives for growing digital biomarkers from any wearable gadget.
