PPG Motion Artifact Removal: Algorithms, Methods & Implementation Guide

Motion artifact removal is the single most important signal processing challenge in photoplethysmography (PPG), and no single algorithm solves it perfectly. The core difficulty is that human motion generates signal distortions in the same 0.1-10 Hz frequency band as the cardiac pulse, making simple bandpass filtering ineffective. This guide covers every major approach to PPG motion artifact removal, from classical adaptive filters to modern deep learning architectures, with practical guidance on selecting the right method for your application.

Whether you are building a wearable heart rate monitor or researching PPG signal processing algorithms, understanding motion artifact removal is essential. The techniques described here form the backbone of every commercial wearable device's heart rate estimation pipeline.

The Motion Artifact Problem in PPG

Photoplethysmography measures blood volume changes by shining LED light into tissue and detecting the reflected or transmitted light with a photodetector. The pulsatile component of this signal (the AC component) corresponds to cardiac-driven blood volume fluctuations and typically has an amplitude of only 1-2% of the total signal. Motion artifacts can easily exceed this amplitude by an order of magnitude.

The fundamental challenge is spectral overlap. At rest, heart rate occupies roughly 0.8-3.3 Hz (48-200 BPM). During walking, the dominant arm swing frequency is 0.8-2.0 Hz. During running, foot strike cadence is 1.3-3.5 Hz. During cycling, pedaling cadence is 1.0-2.0 Hz. These motion frequencies land squarely within the cardiac band, and their harmonics extend the overlap further.

Motion artifacts arise from several physical mechanisms. Sensor-skin coupling changes occur when the device shifts relative to the skin surface, altering the optical path. Tissue deformation from mechanical forces changes local blood distribution. Venous blood pooling from centripetal acceleration during arm swing modulates the optical signal. Ambient light leakage varies with sensor movement. Each mechanism contributes differently depending on the type and intensity of motion, which is why no single removal technique works universally.

For a foundational understanding of how PPG signals are generated and what they contain, see our introduction to PPG technology.

Adaptive Filtering Methods

Adaptive filters are the most widely used class of algorithms for PPG motion artifact removal, primarily because they leverage an accelerometer reference signal to estimate and subtract the motion component. The general framework treats the PPG signal as a mixture of the clean cardiac signal and motion noise. The accelerometer provides a reference correlated with the noise but not with the cardiac signal, allowing an adaptive filter to estimate the noise component.

LMS (Least Mean Squares)

The LMS algorithm is the simplest and most computationally efficient adaptive filter. It updates filter weights using the gradient of the squared error at each sample:

w(n+1) = w(n) + mu * e(n) * x(n)

where mu is the step size, e(n) is the error signal, and x(n) is the reference input. LMS has O(N) complexity per sample, where N is the filter order, making it suitable for resource-constrained embedded systems.

The primary limitation of LMS is its slow convergence, especially when the input signal has a large eigenvalue spread. For PPG applications, where motion characteristics change rapidly (transitioning from walking to running, for instance), LMS may not adapt quickly enough to track the changing artifact. Step size selection is also critical: too large causes instability, too small causes sluggish adaptation.

NLMS (Normalized LMS)

NLMS addresses LMS's convergence issues by normalizing the step size by the input signal power. This makes convergence speed less dependent on signal statistics and more robust to varying motion intensities. NLMS adds minimal computational overhead (one additional division per sample) while substantially improving practical performance.

For real-time wearable applications with moderate motion, NLMS often provides the best tradeoff between complexity and performance. Kim and Yoo (2006) demonstrated effective PPG motion artifact reduction using NLMS with a 3-axis accelerometer reference, achieving heart rate estimation errors below 5 BPM during walking and light jogging.

RLS (Recursive Least Squares)

RLS converges significantly faster than LMS/NLMS by using second-order statistics (autocorrelation matrix estimation) rather than instantaneous gradient. It achieves the optimal Wiener solution at each time step, making it ideal for rapidly changing motion conditions like transitioning between exercise types.

The cost is computational complexity: O(N^2) per sample versus O(N) for LMS. For typical PPG applications with filter orders of 10-50 taps, this remains feasible on modern embedded processors. RLS also has a forgetting factor parameter that controls how quickly it discards old data, which must be tuned for the expected rate of motion change.

RLS consistently outperforms LMS variants in benchmark comparisons but may be overkill for steady-state activities like treadmill running at constant speed. The practical recommendation is to use NLMS as a baseline and upgrade to RLS if performance during activity transitions is insufficient. See our adaptive filtering algorithm guide for implementation details.

Independent Component Analysis (ICA)

ICA takes a fundamentally different approach by treating the PPG and accelerometer signals as mixtures of independent source signals and attempting to recover the original sources through statistical independence maximization. In the PPG context, the goal is to separate the cardiac component from the motion component without explicitly modeling the relationship between them.

The key advantage of ICA is that it does not require a linear noise model or assume a specific relationship between the accelerometer and the PPG motion artifact. It can handle nonlinear mixing to some extent and works even when the accelerometer reference is imperfect. ICA has been shown to effectively separate cardiac and motion components in multi-channel PPG setups (using multiple photodetectors at different locations or wavelengths).

The disadvantages are computational cost (matrix decomposition operations), the requirement for at least as many observation channels as source signals, and the permutation ambiguity problem (determining which recovered component is the cardiac signal). ICA also assumes statistical independence between cardiac and motion signals, which may not hold during highly periodic exercise where cardiac and motion rhythms are coupled. For more on how multi-wavelength PPG can provide additional channels for ICA, see our wavelength guide.

Empirical Mode Decomposition (EMD/EEMD)

Empirical Mode Decomposition is a data-driven method that decomposes a signal into a set of Intrinsic Mode Functions (IMFs) through an iterative sifting process. Each IMF represents an oscillatory component at a characteristic time scale. For motion-corrupted PPG, the idea is that cardiac and motion components will separate into different IMFs based on their frequency and amplitude characteristics.

The standard EMD algorithm suffers from mode mixing, where components of similar frequency end up in the same IMF. Ensemble EMD (EEMD) addresses this by adding white noise to the signal and averaging across multiple decompositions, producing cleaner separation at the cost of increased computation (typically 100-300 ensemble iterations).

EMD/EEMD is attractive because it requires no reference signal and makes minimal assumptions about the signal structure. However, it has several practical limitations. The decomposition is not unique and can be sensitive to signal endpoints and noise levels. Selecting which IMFs correspond to the cardiac signal is non-trivial and typically requires additional heuristic rules or spectral analysis. Computation time for EEMD is substantial, making real-time implementation challenging on low-power wearable processors.

Zhang (2015) incorporated EMD-like spectral decomposition into the TROIKA framework, which combines signal decomposition with sparse signal reconstruction to achieve robust heart rate tracking during exercise. TROIKA demonstrated state-of-the-art performance on the IEEE Signal Processing Cup 2015 dataset and remains a widely cited benchmark method. For related peak detection approaches, our algorithm guide covers complementary techniques.

Wavelet Decomposition

Wavelet transforms decompose the PPG signal into time-frequency components at multiple scales, providing simultaneous time and frequency localization that the Fourier transform cannot achieve. This is particularly useful for PPG motion artifact removal because motion artifacts are often non-stationary (their characteristics change over time).

The discrete wavelet transform (DWT) decomposes the signal into approximation and detail coefficients at each level. Motion artifacts typically appear as high-energy components in specific wavelet scales corresponding to the motion frequency. By thresholding or zeroing the contaminated wavelet coefficients and reconstructing the signal, the motion artifact can be attenuated.

Wavelet-based approaches require selection of the mother wavelet (Daubechies, Symlet, and Coiflet families are common choices for PPG), the decomposition level, and the thresholding strategy (hard vs. soft thresholding, universal vs. level-dependent thresholds). These choices significantly affect performance and are often determined empirically. The continuous wavelet transform (CWT) provides finer frequency resolution but at higher computational cost.

Wavelet methods work well for impulsive or transient motion artifacts (sudden arm movements, tapping) but are less effective for sustained periodic motion (running, cycling) where the artifact is stationary and spectrally overlapping with the cardiac signal across all time windows.

Singular Value Decomposition (SVD)

SVD-based methods construct a Hankel matrix from the PPG signal and decompose it into singular value components. The dominant singular values typically correspond to the strongest signal components (motion artifact during heavy exercise), while smaller singular values capture the cardiac signal. By selectively reconstructing the signal from appropriate singular value components, motion artifacts can be attenuated.

SVD is computationally expensive (O(N^3) for an N-length segment) and is generally used in offline processing rather than real-time applications. It is most effective when combined with other methods as part of a multi-stage pipeline, where SVD handles the initial gross artifact removal and subsequent stages refine the cardiac signal extraction. This approach works well for blood oxygen estimation where signal quality directly impacts SpO2 accuracy.

Deep Learning Approaches

Deep learning has transformed PPG motion artifact removal over the past several years, consistently achieving the lowest error rates on benchmark datasets. Several architectures have proven effective.

Convolutional Neural Networks (CNNs)

1D CNNs process fixed-length PPG windows and learn hierarchical features that capture both local waveform morphology and broader spectral patterns. CNNs excel at learning the mapping from corrupted PPG segments to clean cardiac signals or directly to heart rate estimates. They are relatively efficient for inference and can run in real-time on mobile processors.

LSTM and Recurrent Networks

Long Short-Term Memory networks capture temporal dependencies across PPG windows, making them effective for tracking heart rate through motion-corrupted segments where the cardiac signal temporarily disappears. Bidirectional LSTMs can look both forward and backward in time, improving accuracy for offline processing.

U-Net Architectures

U-Net and similar encoder-decoder architectures with skip connections have shown strong performance for PPG denoising as a signal-to-signal translation task. The encoder compresses the corrupted signal into a latent representation, and the decoder reconstructs the clean signal, with skip connections preserving fine-grained temporal details. Mashhadi et al. (2021) demonstrated that U-Net based approaches achieve competitive or superior performance compared to traditional methods on standard benchmarks.

Practical Considerations for Deep Learning

Deep learning methods require substantial training data with ground-truth labels (typically ECG-derived heart rate). They can overfit to specific motion patterns or sensor hardware if training data is not diverse. Model size and inference latency must be carefully managed for embedded deployment. Transfer learning from large PPG datasets can help when application-specific data is limited. For a broader view of how deep learning integrates with wearable sensor technology, see our overview.

The Complete Processing Pipeline

A practical PPG motion artifact removal pipeline typically includes these stages in sequence:

1. Pre-processing. Band-pass filtering (0.4-4 Hz for heart rate applications) to remove DC offset, high-frequency noise, and powerline interference. Accelerometer data is similarly filtered.

2. Signal quality assessment. Evaluate the PPG signal quality index (SQI) to determine whether motion artifact removal is needed and to flag segments that are too corrupted to recover.

3. Gross artifact removal. Adaptive filtering (NLMS or RLS) using the accelerometer reference to remove the primary motion component.

4. Fine artifact removal. Spectral analysis (STFT or wavelet) combined with spectral peak tracking to identify and extract the cardiac frequency component from residual noise.

5. Heart rate estimation. Spectral peak selection with physiological constraints (maximum rate of change, harmonic rejection) to produce the final heart rate estimate.

6. Post-processing. Median filtering or Kalman filtering of the heart rate time series to smooth outliers and fill gaps from unrecoverable segments.

This multi-stage approach is the architecture used by most commercial wearables, with specific algorithms varying by manufacturer. The TROIKA framework (Zhang, 2015) and its successor JOSS follow this general pattern with specific spectral sparse signal recovery techniques at the core.

Benchmark Datasets and Evaluation

Two datasets dominate the PPG motion artifact removal literature. The IEEE Signal Processing Cup 2015 dataset contains 12 recordings of wrist PPG with simultaneous ECG reference during treadmill running at speeds from 1-15 km/h. It remains the most widely used benchmark for comparing algorithms, with state-of-the-art methods achieving mean absolute error below 2 BPM on this dataset.

The PPG-DaLiA dataset (Reiss et al., 2019) is a larger and more diverse benchmark with 15 subjects performing eight different daily-life activities including sitting, walking, cycling, driving, and playing table tennis. It provides wrist PPG, chest ECG, and 3-axis accelerometer data, making it suitable for evaluating algorithms across a range of real-world scenarios rather than just treadmill running.

When evaluating motion artifact removal algorithms, the key metrics are mean absolute error (MAE) and mean absolute percentage error (MAPE) of heart rate estimation against the ECG reference. Coverage (percentage of time windows where a valid estimate is produced) is also important, as some algorithms achieve low error by simply refusing to estimate during highly corrupted segments. For those building custom evaluation pipelines, our PPG analysis tools guide provides practical resources.

Real-Time Implementation Considerations

For embedded wearable deployment, algorithm selection is constrained by processor speed, memory, power consumption, and latency requirements. Adaptive filters (NLMS, RLS) are the most practical for low-power microcontrollers, requiring only kilobytes of RAM and microseconds per sample. Wavelet and EMD methods are feasible on application processors (ARM Cortex-A class) but may be too expensive for sensor-hub processors. Deep learning inference requires careful model optimization (quantization, pruning) to fit within wearable power budgets.

Latency is another critical factor. Most methods process data in windows of 4-8 seconds with 1-2 second overlap, introducing 2-4 seconds of algorithmic latency on top of any buffering delays. For exercise applications this is acceptable, but for real-time biofeedback or clinical monitoring, lower-latency approaches may be needed. Understanding these tradeoffs is essential whether you are developing for a smartwatch platform or a clinical device.

Frequently Asked Questions

What is the best algorithm for PPG motion artifact removal?

There is no single best algorithm for all scenarios. Adaptive filtering methods like NLMS and RLS work well when a clean accelerometer reference is available and computational resources are limited. For offline or high-accuracy applications, deep learning approaches (CNN-LSTM hybrids, U-Net architectures) consistently achieve the lowest heart rate estimation error on benchmark datasets like IEEE SP Cup 2015. The best choice depends on your hardware constraints, latency tolerance, and whether a motion reference signal is available.

Can motion artifacts be completely removed from PPG signals?

Complete removal is not possible in all cases. When motion artifact frequency components fully overlap with the cardiac signal (both in the 0.5-4 Hz range during vigorous exercise), some information is irrecoverably lost. However, modern multi-stage pipelines combining accelerometer-referenced adaptive filtering with spectral tracking can recover usable heart rate estimates in roughly 85-95% of motion-corrupted segments, depending on the intensity and type of motion.

Why do PPG sensors fail during intense exercise?

During intense exercise, several factors degrade PPG signal quality simultaneously. Vigorous limb movement causes the sensor to shift on the skin, creating optical coupling variations. Increased blood flow and vasodilation change the baseline signal amplitude. Most critically, the rhythmic motion of running or cycling generates periodic artifacts whose frequencies (1-4 Hz at typical cadences) overlap directly with the heart rate frequency band, making it extremely difficult to separate cardiac from motion components using simple filtering.

Do I need an accelerometer for PPG motion artifact removal?

An accelerometer is not strictly required but significantly improves results. Methods like ICA, EMD, and some deep learning approaches can operate on the PPG signal alone (single-channel). However, having a co-located 3-axis accelerometer provides a direct reference of the motion corrupting the signal, enabling adaptive filtering methods that are both computationally efficient and highly effective. Nearly all modern wearable devices include accelerometers, making reference-based methods the practical standard.