Abstract
Continuously operating wearables offer detailed insight into chronic health conditions and have the potential to reshape diagnostic and screening tools. However, the energy demands and large datasets created by constant monitoring over weeks to months are difficult or impossible to integrate into existing clinical practice, limiting the utility of this device class. Machine learning offers the opportunity to condense these large datasets into streamlined, digestible trends with the potential for significant clinical impact, although off-device inference requires advanced network infrastructure and substantial power availability for radios. Here, we introduce a device framework that integrates artificial intelligence with clinical grade biosignal acquisition at the edge, performing on-device inference with clinical grade fidelity over extended durations with no interaction required by the wearer. We utilize this framework to perform gait-based frailty assessment during in vivo trials (N(1) = 16) with results that match gold standard diagnostic tools. Clinical utility, model stability, and on-device inference are validated through in vivo trials (N(2) = 14) and ten-day-long extended wear experiments, demonstrating continuous operation without wearer intervention and autonomous longitudinal analysis of high sampling rate biosignals.