ChatPPG Editorial

Wavelet Denoising for PPG Signals

Wavelet denoising is one of the most effective reference-free techniques for cleaning noisy PPG because it separates transient artifact, baseline drift, and physiologic structure across multiple scales instead of relying on a single fixed cutoff.

ChatPPG Research Team
8 min read
Wavelet Denoising for PPG Signals

Wavelet denoising is one of the most effective reference-free techniques for cleaning noisy PPG because it separates transient artifact, baseline drift, and physiologic structure across multiple scales instead of relying on a single fixed cutoff. That matters in photoplethysmography because noise is rarely confined to one band. The same recording can contain slow baseline drift, motion bursts, LED shot noise, and high-frequency sensor contamination. Wavelets offer a framework for handling all of them in a scale-aware way.

If you have already read our article on PPG wavelet transform denoising, think of this piece as the implementation-first companion focused on wearable signal cleaning strategy. It also pairs well with our guides on baseline wander removal and noise reduction techniques for PPG.

Why wavelets are attractive for PPG

Fourier filtering treats signal content as globally stationary. PPG noise is not. A motion burst during one second of a wrist recording does not mean the whole clip is noisy. Wavelets handle this better because they localize structure in both time and scale. That lets the engineer suppress coefficients likely to represent noise while preserving coefficients likely to represent physiologic pulse morphology.

For PPG, that is especially useful because the waveform contains a steep systolic upstroke, rounded systolic peak, and slower diastolic decay. Those features occupy different scales, and their relative prominence changes with age, vasomotor tone, sensor site, and sampling rate.

Discrete wavelet transform basics

The discrete wavelet transform, or DWT, decomposes the signal into approximation and detail coefficients using paired low-pass and high-pass filters followed by downsampling. After several levels of decomposition, the signal is represented as a hierarchy of scales.

In practical PPG terms:

  • high-level approximation coefficients often capture baseline drift and very slow modulation
  • mid-level detail coefficients often capture cardiac structure
  • low-level detail coefficients often capture sensor noise and sharp artifact

The precise frequency mapping depends on sample rate. At 100 Hz, the first few detail levels cover very different bands than they do at 25 Hz. This is why decomposition depth should never be copied blindly from a paper without checking the device sample rate.

Mallat's multiresolution formulation underpins this entire approach and remains a foundational citation in engineering work:

Choosing a mother wavelet for PPG

One of the most common questions is which wavelet family to use. In PPG studies, Daubechies and Symlets are common because their compact support and shape often align reasonably well with pulse morphology. Common practical choices include db4, db6, sym4, and coif variants.

Selection criteria should include:

  1. similarity to the pulse shape
  2. acceptable phase behavior for downstream timing tasks
  3. reconstruction quality after thresholding
  4. computational simplicity for embedded deployment

There is no universally best wavelet. A wrist PPG heart-rate application and a finger PPG morphology study can favor different choices. If you are mainly extracting heart rate, preserving gross cardiac energy may be enough. If you need derivative landmarks or dicrotic notch characterization, waveform preservation deserves more weight.

Thresholding strategy matters more than many teams realize

Wavelet denoising is not just about decomposition. The thresholding rule determines whether the method behaves like a careful cleaner or a waveform destroyer.

The classical workflow is:

  1. decompose the signal
  2. threshold selected detail coefficients
  3. reconstruct the signal

Hard thresholding

Hard thresholding sets coefficients below a threshold to zero and keeps the rest unchanged. It preserves large coefficients well, but it can create discontinuities at the threshold boundary and may introduce ringing in the reconstructed waveform.

Soft thresholding

Soft thresholding shrinks coefficients toward zero. The reconstruction is smoother and often visually cleaner, but it can attenuate physiologic detail if the threshold is too aggressive.

In PPG, soft thresholding is often preferred for general denoising because it reduces choppy artifacts. Hard thresholding may preserve pulse amplitude slightly better in some settings, though the result can be less smooth.

Threshold selection

Several threshold rules are common:

  • universal threshold
  • SURE or Stein-based thresholding
  • BayesShrink
  • level-dependent empirical thresholds

The universal threshold is easy but often too aggressive for biologic waveforms. BayesShrink and level-dependent approaches are usually better starting points for PPG because different decomposition levels have different signal-to-noise characteristics.

Donoho and Johnstone established much of the foundational thresholding literature:

How wavelet denoising maps to common PPG noise sources

Baseline wander

Very slow drift from respiration, vasomotor modulation, temperature shifts, or contact-pressure change tends to fall into coarse approximation coefficients. Removing or attenuating only those low-frequency components can flatten the baseline without applying an aggressive high-pass filter to the entire signal. For a deeper dive, see our PPG baseline wander removal guide.

High-frequency sensor noise

Electronic noise and quantization effects often sit in the finest detail coefficients. These are usually the easiest coefficients to suppress because they tend to contribute little to pulse morphology in most wearable sample-rate regimes.

Motion bursts

Motion artifact is trickier because it may span multiple scales and can overlap with the cardiac band. Wavelet denoising alone can help when the artifact is transient or broadband, but periodic motion that overlaps heart rate may require augmentation with adaptive filtering or Kalman tracking.

Level selection and sample rate awareness

A repeated mistake in PPG implementations is choosing a decomposition level without thinking through the actual frequency mapping.

Suppose the sample rate is 128 Hz. A six-level decomposition places different physiologic components into different subbands than a six-level decomposition at 25 Hz. The heart-rate band, usually about 0.7 to 3.5 Hz in many adult monitoring contexts, can sit partly in coarse detail levels and partly in approximation depending on sampling.

Because of this, teams should:

  • map expected physiologic bands to wavelet levels explicitly
  • verify with synthetic signals before deploying on real recordings
  • inspect whether reconstruction preserves systolic timing and pulse amplitude

This step is just as important as selecting the mother wavelet.

Wavelet denoising for morphology versus heart rate

Not every PPG application values the same output.

Heart rate estimation

If the end goal is stable heart rate, then preserving dominant periodic energy is the main objective. A denoised waveform that is slightly smoothed may still perform very well if peak timing remains stable or the frequency spectrum is cleaner.

Morphology analysis

If the end goal is arterial stiffness markers, notch timing, augmentation index, or derivative features, then smoothing can become dangerous. Small shape distortions can create large errors in derived markers. In these cases, the denoising settings should be validated against morphology-specific metrics, not just output SNR.

This distinction is central to good PPG engineering. A denoiser that looks clean may still be the wrong tool if it removes clinically relevant fine structure.

Stationary wavelet transform versus DWT

The stationary wavelet transform, or SWT, avoids downsampling and therefore maintains shift invariance. That can be valuable in PPG because pulse timing should not depend on arbitrary alignment to the decomposition grid.

The tradeoff is cost. SWT uses more memory and more computation than the standard DWT. For offline analysis or server-side processing, that may be acceptable. For battery-constrained wearables, DWT is often more realistic.

If you are evaluating beat-to-beat timing or pulse transit surrogates, the extra timing stability of SWT may justify the overhead.

Hybrid pipelines often outperform wavelets alone

Wavelet denoising is strong, but the best real-world pipelines often combine methods.

Examples include:

  • broad bandpass filtering followed by wavelet denoising
  • wavelet denoising followed by robust peak detection algorithms
  • adaptive filtering with accelerometer reference plus wavelet cleanup
  • signal quality gating before wavelet reconstruction

This layered approach works because no single method handles every noise source equally well. Wavelets are particularly valuable as a reference-free denoising core inside a broader processing stack.

Evaluation metrics that matter

Do not evaluate wavelet denoising using only subjective waveform appearance. Better metrics include:

  • heart-rate MAE versus ECG
  • beat detection sensitivity and positive predictive value
  • normalized root mean square error on synthetic benchmarks
  • correlation of morphology landmarks
  • output SNR improvement under controlled noise injection

Synthetic benchmarks are helpful because you know the clean reference. Real wearable datasets are still essential because synthetic noise rarely captures all the quirks of real sensor-skin coupling.

Common implementation mistakes

Teams often struggle for the same reasons:

  1. choosing decomposition levels without frequency mapping
  2. using thresholds that are too aggressive for morphology work
  3. assuming the same settings work across wrist, finger, and ear PPG
  4. validating on resting data only
  5. judging success from visual smoothness instead of endpoint metrics

A careful engineer treats wavelet denoising as a tunable signal model, not a checkbox preprocessing step.

Practical recommendations

If you are implementing wavelet denoising for PPG today:

  • start with db4 or sym4
  • use level-dependent soft thresholding rather than one global threshold
  • map decomposition levels to the device sample rate explicitly
  • test on both clean and motion-contaminated segments
  • compare results separately for heart-rate and morphology use cases

Wavelet denoising remains one of the best reference-free options for PPG because it gives fine-grained control across scales. Its strength is flexibility, but that same flexibility demands disciplined tuning and validation.

FAQ

What is wavelet denoising in PPG?

Wavelet denoising decomposes the PPG signal into multiple scales, suppresses coefficients likely to represent noise, and reconstructs the signal. It is useful because PPG noise often spans several frequencies and changes over time.

Is wavelet denoising better than a bandpass filter for PPG?

For many noisy wearable signals, yes, especially when the noise is transient or multi-scale. A bandpass filter is simpler and still useful, but wavelets can preserve important physiologic content more selectively when tuned well.

Which wavelet is best for PPG signals?

There is no universal winner. Daubechies and Symlet families, especially db4 and sym4, are common starting points. The best choice depends on sample rate, sensor site, and whether the goal is heart rate or morphology preservation.

Does wavelet denoising distort the PPG waveform?

It can if the thresholding is too aggressive or the wavelet and decomposition levels are poorly chosen. That is why validation should include morphology and timing metrics, not just visual inspection.

Can wavelets remove motion artifact from PPG?

They can reduce some motion artifact, especially transient broadband contamination. Strong periodic motion that overlaps the heart-rate band may still require adaptive filtering, Kalman methods, or quality gating.

Is wavelet denoising practical for embedded wearables?

Yes, in many cases. Standard DWT implementations are computationally manageable on modern embedded hardware, though SWT and more complex hybrids may be better suited to phones, gateways, or offline processing.

References

← Back to all articles