Wrist fitness trackers have a well-known problem: move your arm and the readings go sideways.
Researchers have proposed a framework called Physically-Constrained Harmonic Separation, or PCHS, that takes a different approach to cleaning up wrist photoplethysmography data. Instead of training a neural network to directly guess your heart rate or respiratory rate, PCHS treats the problem as analysis-by-synthesis: it uses accelerometer data to model motion artifacts separately, then decomposes the raw optical signal into physiological components and a motion-related residual. Heart rate comes from the fundamental frequency of the physiological component; respiratory rate comes from subtler modulations in that signal's harmonic structure. Tested on the PPG-DaLiA dataset — a benchmark specifically built for high-motion scenarios — PCHS outperformed existing methods while producing decompositions that can actually be inspected and understood.
The gap between lab-grade pulse oximeters and consumer wearables has always been motion: optical sensors on the wrist pick up every arm swing as noise that sits in the same frequency range as a beating heart. Most deep learning fixes for this work well on clean data and fall apart in the wild, partly because they offer no way to audit what the model learned. A physics-guided approach that separates signals by their known harmonic structure is harder to fool and easier to trust.
The research does not yet represent a shipping product — it is a preprint, and benchmarks on controlled datasets do not guarantee real-world accuracy across diverse populations and wrist placements. But it is the kind of foundational work that could inform the next generation of health-tracking algorithms inside devices from Apple, Garmin, or Fitbit, all of which face the same motion-artifact problem and have not fully solved it.