Robust Physiological Monitoring Under Severe Motion Artifacts

1,202 concepts796 papers11 proposals

Continuous physiological monitoring—whether tracking a patient's heart rate via a smartwatch or measuring respiration rates remotely using radar—promises to revolutionize healthcare. However, moving these technologies from controlled laboratory environments to real-world deployment introduces a fundamental physical challenge: the scale disparity problem.

Physiological signals are microscopic. A heartbeat causes sub-millimeter skin color shifts (via blood volume changes) or minute wrist vibrations. Respiration produces millimeter-scale chest displacements. Conversely, everyday human activities like walking, talking, or turning over involve macroscopic movements measured in centimeters or meters. When a subject moves, these macro-movements generate signal shifts that are orders of magnitude larger than the target physiological data, creating severe motion artifacts (MAs) that completely mask the underlying vital signs.

To unlock continuous, "in-the-wild" health tracking, researchers are developing signal processing and machine learning frameworks designed to mathematically isolate micro-physiological signals from overwhelming macro-motion noise.

The Nature of Motion Interference

To understand why motion artifacts are so destructive, consider how non-contact and wearable sensors capture data. In wireless sensing (such as Wi-Fi or radar), radio waves bounce off the body to detect the rhythmic movement of the chest. As noted by Luzhou Xu et al. in Soli-enabled noncontact heart rate detection for sleep and meditation tracking, the mechanical displacement of a heartbeat is dwarfed by respiration, which is itself dwarfed by gross body movements.

When a person walks or adjusts their posture, the reflections create chaotic signal fluctuations. Xiang Zhang et al. demonstrate in Wital: A COTS WiFi Devices Based Vital Signs Monitoring System Using NLOS Sensing Model that these large-scale irregular motions act as a massive DC offset, drowning out the low-amplitude periodic signals of interest.

A similar vulnerability plagues optical sensors like photplethysmography (PPG), which smartwatches use to track heart rates. Physical activity introduces extreme noise by shifting the sensor relative to the skin or altering local blood flow dynamics. As described by Yichen Shen 0003 et al. in Ambulatory Atrial Fibrillation Monitoring Using Wearable Photoplethysmography with Deep Learning, this interference frequently causes standard algorithms to lose the pulse signal entirely, leading to dangerous misclassifications or forcing systems to discard large segments of data.

Traditional frequency filtering fails to resolve this because the frequency of human motion (e.g., a walking stride rate) often directly overlaps with the frequencies of human heart rates and respiration.

Disentangling Motion via Deep Learning

Rather than relying on rigid frequency filters, modern approaches use deep neural networks to learn the structural differences between chaotic macro-movements and periodic micro-movements.

One powerful intuition is to use adversarial learning to strip motion features away from the raw data. In Remote Respiration Monitoring of Moving Person Using Radio Signals, Jae-Ho Choi et al. combined Doppler-based signal processing with a multi-task adversarial framework. By training a discriminator network to identify and penalize motion-specific features, the feature extractor is forced to ignore global body translations (like walking in a room) and isolate the residual, millimeter-scale periodic chest displacements of breathing.

Raw Signal (Vitals + Motion) ──> [ Feature Extractor ] ──┬──> [ Respiration Predictor ]
                                                         │
                                                         └──> [ Motion Discriminator ] (Penalizes motion features)

Another strategy leverages generative latent spaces to separate signal components. In MoRe-Fi: Motion-robust and Fine-grained Respiration Monitoring via Deep-Learning UWB Radar, Chao Cai 0001 et al. utilized a Variational Encoder-Decoder network to process complex radar signals. The network maps the incoming data into a lower-dimensional latent space where the high-amplitude, irregular variations of physical exercise can be mathematically disentangled from fine-grained physiological waveforms.

For video-based remote physiological measurement (rPPG), where head tilt and facial expressions distort the pixel values, unsupervised learning can find the underlying signal. In Contrast-Phys: Unsupervised Video-Based Remote Physiological Measurement via Spatiotemporal Contrast, Xiaobai Li introduced a spatiotemporal contrastive loss. The core insight is that while motion artifacts vary wildly across different parts of the face and time windows, the underlying blood volume pulse remains uniform and synchronized across all skin regions. By forcing the network to maximize the similarity of extracted signals across different facial zones, it naturally learns to ignore localized, non-periodic motion noise.

Multi-Modal and Multi-Wavelength Anchoring

When a single data stream is corrupted, adding a secondary physical dimension can provide a mathematical anchor to calculate and subtract the noise.

The most direct implementation of this is multi-sensor fusion. In Physiological signal monitoring in the waiting areas of an emergency room, Dorothy Curtis et al. demonstrated that integrating secondary sensors like accelerometers alongside primary physiological sensors (like ECG) allows algorithms to explicitly measure physical acceleration (e.g., arm swinging). This motion profile is then used as a reference to differentiate between actual pathological cardiac changes and motion-induced signal artifacts.

In optical wearable sensing, this concept manifests as multi-wavelength fusion. Photons of different colors penetrate human tissue to different depths; green light is heavily absorbed by superficial blood vessels, while red and infrared light penetrate deeper. Crucially, motion artifacts affect these optical channels globally, whereas the blood volume pulse affects them differentially based on tissue depth and hemoglobin absorption spectra.

In Tri-Spectral PPG: Robust Reflective Photoplethysmography by Fusing Multiple Wavelengths for Cardiac Monitoring, Manuel Meier et al. leveraged this spectral redundancy. By feeding green, red, and infrared channels simultaneously into a U-Net architecture, the model exploits the shared motion noise across channels to isolate and reconstruct a clean, uncorrupted cardiac waveform.

Similarly, Jehyun Kyung et al. showed in Deep-learning-based blood pressure estimation using multi channel photoplethysmogram and finger pressure with attention mechanism that capturing signals from multiple spatial channels simultaneously allows attention mechanisms to dynamically de-emphasize channels suffering from localized motion or poor skin contact.

Generalizing to the Wild

Historically, motion-compensation algorithms were brittle because they were trained on small, highly controlled laboratory datasets featuring scripted, rhythmic movements. Real-world motion, however, is chaotic and highly variable. Furthermore, factors like skin tone diversity alter how light interacts with tissue, compounding the difficulty of separating pulse signals from noise.

To force models to learn truly robust representations, the research community has shifted toward diverse benchmarks and large-scale pre-training. Jiankai Tang et al. released MMPD: Multi-Domain Mobile Video Physiology Dataset, a mobile video dataset explicitly capturing varied motion, lighting, and darker skin tones. This benchmark revealed that models optimized for clean lab data degrade significantly under real-world conditions, highlighting the need for generalized feature extraction.

The ultimate realization of this trend is the development of foundation models. In PaPaGei: Open Foundation Models for Optical Physiological Signals, Arvind Pillai et al. introduced a morphology-aware foundation model pre-trained on over 57,000 hours of physiological data. By learning the fundamental, intrinsic grammar of physiological waveforms across thousands of individuals and environments, these models develop an internal resilience to noise. They recognize what a true physiological signal should look like, allowing them to reliably track health metrics through severe, real-world motion interference where traditional algorithms fail.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_motion_artifact_interference_6d81c7e6,
  title = {motion artifact interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/motion-artifact-interference/rdr_6d81c7e605c64c78894b0949f957000d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_motion_artifact_interference_6d81c7e6_v1,
  title = {motion artifact interference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/motion-artifact-interference/rdr_6d81c7e605c64c78894b0949f957000d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake