Paper Overview
- Field: Health AI
- Authors: Tanbin Islam Rohan, Pranjol Sen Gupta, Tanusree Debi, Nazmus Sakib
- Published: 2026-10-05
- arXiv: 2610.06823
- Spectral descriptors
- An intermediate Nightbeat-derived frequency anchor
- A physics-inspired subharmonic feature designed to identify harmonic frequency locking
- SeqSmoother: 1.60 bpm participant-macro MAE
- Error on Nightbeat-retained matched intervals: Nightbeat achieved lower absolute error (0.615 vs 1.091 bpm)
- Coverage: Nightbeat produced final estimates for only 72.85% of SeqSmoother's eligible out-of-sample grids—SeqSmoother delivered estimates on a substantially larger share of eligible records
- Subharmonic ratio: AUROC of 0.972 for identifying reference-defined harmonic-locking candidates
Summary
Large longitudinal cohorts often contain wrist accelerometer data but lack optical heart rate sensing, motivating the recovery of cardiac information from motion signals already collected during sleep. The authors propose SeqSmoother, a transformer-based temporal corrector for sleep heart rate (HR) estimation.
SeqSmoother combines:
All inference-time features are derived from the wrist accelerometer, while ECG is used only to construct reference HR labels and training label-quality weights.
Results
Evaluation used held-out folds separated by participant (13 participants), compared against Nightbeat's official implementation under a matched 60-second window and 15-second step protocol:
Takeaway
These findings reveal a precision-availability trade-off between learned temporal modeling and quality-gated signal processing: a quality-gated baseline can be more accurate where it reports results, but learned correction offers far broader coverage at competitive overall accuracy.
Original Abstract
SeqSmoother is a transformer-based temporal corrector for sleep heart rate estimation from wrist accelerometry, achieving 1.60 bpm participant-macro MAE with broader coverage than the quality-gated Nightbeat baseline.