Photoplethysmography in wearable devices is highly vulnerable to motion artifacts, disturbances caused when sensor movement, ambient light, or tissue deformation obscure the pulsatile blood-volume signal. These artifacts matter because they degrade heart rate and heart rate variability estimates used in clinical screening, fitness tracking, and epidemiological studies. Mohamed Elgendi University of British Columbia Okanagan has characterized PPG morphology and emphasized the need for preprocessing to preserve clinically relevant features. Sources of corruption vary by wear location, skin tone, and activity type, so no single fix is universally optimal.
Adaptive filtering and sensor fusion
Adaptive algorithms such as least-mean-squares and recursive least squares remove correlated motion noise when an auxiliary reference is available. Using a tri-axial accelerometer as the reference signal allows the adaptive filter to track changing motion patterns and subtract correlated components in real time. Gary D. Clifford University of Oxford has demonstrated that sensor fusion combining accelerometry and PPG substantially improves beat detection during ambulatory activities. Adaptive approaches require careful selection of step sizes and may struggle when motion and cardiac components overlap spectrally.
Frequency-domain, wavelet, and decomposition methods
Bandpass filtering isolates the typical cardiac band while rejecting low-frequency baseline wander and high-frequency jitter, but simple filtering can distort pulse shape. Wavelet denoising and empirical mode decomposition separate oscillatory modes, enabling targeted suppression of motion-related modes while retaining systolic peaks. Independent component analysis can also separate physiological sources in multisensor setups. More recently, supervised machine learning and deep-learning models trained on labeled motion-corrupted PPG can learn mappings from noisy to clean signals, improving robustness during complex activities. Machine learning methods improve performance but require diverse training data to avoid bias.
Motion artifact mitigation has human, cultural, and environmental implications. Device performance often varies across skin pigmentation and wristwear practices, influencing the equity of health monitoring in different populations. Occupational or cultural patterns of movement change artifact profiles, and environmental temperature affects peripheral perfusion and PPG amplitude. Consequences of inadequate artifact reduction include false alarms, missed arrhythmias, and unreliable longitudinal metrics that can misinform users and clinicians.
Combining adaptive filtering, accelerometer fusion, decomposition techniques, and judicious machine learning typically yields the best trade-off between accuracy and real-time constraints. Designers must balance computational load and battery life against signal fidelity, validate across diverse populations, and report methods transparently to support clinical and public trust. Residual errors remain a practical limitation for some high-motion scenarios and for metrics that rely on precise waveform morphology.