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

Large longitudinal cohorts often contain wrist accelerometry without optical heart-rate sensing, motivating recovery of cardiac information from motion signals already collected during sleep. We present SeqSmoother, a transformer-based temporal corrector for sleep heart rate HR estimation from wrist accelerometry. 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. All inference-time features are derived from wrist accelerometry, while ECG is used only to construct reference HR labels and training-label quality weights. We evaluate SeqSmoother using 13 participant-disjoint held-out folds and compare it with the official Nightbeat implementation under a matched 60-s window and 15-s step protocol. Across all out-of-fold predictions, SeqSmoother achieved a participant-macro MAE of 1.60 bpm. On Nightbeat-retained matched intervals, Nightbeat achieved lower absolute error than SeqSmoother 0.615 versus 1.091 bpm, while SeqSmoother provided estimates over a larger portion of the eligible recording; Nightbeat produced final estimates for 72.85% of the SeqSmoother-eligible out-of-fold grid. Separately, the proposed sub-harmonic ratio achieved an AUROC of 0.972 for identifying reference-defined harmonic lock-on candidates. These findings reveal an accuracy-availability trade-off between learned temporal modeling and quality-gated signal processing while providing empirical support for a physics-informed approach to identifying frequency-tracking failures in accelerometer-based sleep HR estimation.

Sleep heart rate recovery from wrist-worn motion

Population-scale health studies frequently deploy wrist-worn devices to monitor activity and sleep. Many of these devices lack optical heart-rate sensors, relying instead on tri-axial accelerometry. While accelerometers capture motion, recovering physiologically meaningful cardiac signals from pure motion data has remained challenging. This paper addresses the gap by proposing SeqSmoother, a learning-based corrector that estimates sleep heart rate from wrist accelerometry alone. The work is motivated by the reality that large existing cohorts contain only motion streams, and any method that can extract cardiac insight from those streams increases the value of the data without requiring new hardware.

SeqSmoother architecture

SeqSmoother operates as a transformer-based temporal corrector. The model takes as input spectral descriptors computed from wrist accelerometry during sleep. In addition, the model uses an intermediate frequency anchor derived from the Nightbeat reference method, and a physics-motivated sub-harmonic feature. The sub-harmonic feature is designed to identify when the observed frequency lock-on to a harmonic of the true heart rate occurs. All features at inference time are derived from the accelerometer signal; electrocardiogram ECG is used only during label construction and for assigning training-weight values. This separation means the model can be deployed in settings where only motion data are available.

The transformer processes the sequence of spectral descriptors and outputs a heart-rate estimate for each time window. The frequency anchor provides a starting point derived from the established Nightbeat approach, while the sub-harmonic feature adds a physics-informed dimension that helps the model detect when frequency tracking may have failed.

Experimental protocol

Evaluation uses 13 participant-disjoint held-out folds. That means the model is tested on participants who were not seen during training, ensuring that results generalize across individuals. The comparison runs against the official Nightbeat implementation under a matched protocol of 60-second windows with 15-second steps. This protocol ensures that both methods operate on the same temporal granularity, making the comparison fair.

Across all out-of-fold predictions, SeqSmoother achieved a participant-macro mean absolute error of 1.60 beats per minute. On the intervals where Nightbeat produced a final estimate, Nightbeat recorded lower absolute error: 0.615 bpm versus 1.091 bpm for SeqSmoother. However, SeqSmoother produced estimates over a larger portion of the eligible recording. Nightbeat provided final estimates for 72.85% of the grid that SeqSmoother marked as eligible. This contrast between accuracy and coverage forms the central trade-off explored in the paper.

Sub-harmonic ratio for frequency-tracking assessment

Beyond the point-estimate error, the paper introduces a sub-harmonic ratio feature. This feature is computed from the accelerometer signal and is designed to indicate whether the observed heart-rate frequency has locked on to a sub-harmonic of the true rate. The authors evaluate this feature against reference-defined harmonic lock-on candidates, achieving an area under the receiver operating characteristic curve of 0.972. An AUROC this high indicates that the sub-harmonic ratio reliably distinguishes cases where the accelerometer-based tracking has converged to an incorrect harmonic versus cases where the tracking is accurate.

Accuracy-availability trade-off

The results reveal a consistent pattern. Nightbeat, which relies on quality-gated signal processing, achieves lower absolute error on the subset of the recording it can process. SeqSmoother, which uses learned temporal modeling across the full accelerometer stream, covers more of the recording but with higher error on individual windows. This accuracy-availability trade-off suggests that neither approach dominates entirely; the choice between them may depend on whether the application prioritizes per-window accuracy or overall recording coverage.

The authors note that the physics-informed sub-harmonic feature provides empirical support for identifying frequency-tracking failures. When the sub-harmonic ratio indicates a lock-on to an incorrect harmonic, the estimated heart rate can be substantially off. This finding points toward a combined approach: use the sub-harmonic ratio to flag potentially unreliable windows, and apply the model selectively or apply a correction factor.

Population-scale implications

Making cardiac insight available from existing accelerometer data has implications for the study of large cohorts. Researchers studying aging, cardiovascular health, or sleep disorders in population-scale datasets can obtain heart-rate information without deploying additional sensors. The trade-off identified in this work—between the precision of quality-gated methods and the coverage of learned methods—informs how such researchers might design their analysis pipelines.

The paper contributes a practical method for extracting cardiac signals from motion data that is already being collected at scale. By releasing the sub-harmonic ratio feature and the experimental protocol, the authors enable other researchers to evaluate the method on their own cohorts and to explore hybrid approaches that combine the strengths of both quality-gated and learned techniques.