AI TrackerFrontier desk · AEST
Back to news
normalaitech

Deep Learning for Sleep Heart Rate Estimation from Accelerometers: Toward Population-Scale Cardiac Insight Without Optical Sensors

arXiv cs.AI · 6 Oct 2026, 03:56 AEST

A new deep learning model, SeqSmoother, has been developed for estimating sleep heart rates from wrist accelerometry data. This model aims to enhance cardiac insights without the need for optical sensors, leveraging existing motion signals collected during sleep.

Key facts

Read aloud

A new deep learning model, SeqSmoother, has been developed for estimating sleep heart rates from wrist accelerometry data. This model aims to enhance cardiac insights without the need for optical sensors, leveraging existing motion signals collected during sleep. SeqSmoother combines spectral descriptors with an intermediate Nightbeat-derived frequency anchor and a physics-motivated sub-harmonic feature designed to identify harmonic frequency lock-on. Across all out-of-fold predictions, SeqSmoother achieved a participant-macro MAE of one point six zero bpm. On Nightbeat-retained matched intervals, Nightbeat achieved lower absolute error than SeqSmoother zero point six one five versus one point zero nine one bpm, while SeqSmoother provided estimates over a larger portion of the eligible recording. The proposed sub-harmonic ratio achieved an AUROC of zero point nine seven two for identifying reference-defined harmonic lock-on candidates.

Sources in this cluster