Bandpass Filter for PPG, Practical Design for Pulse Retention

A bandpass filter is often the first serious processing step in a PPG pipeline, but good design is more than picking 0.5 to 5 Hz and moving on. The right passband depends on sample rate, measurement site, target physiology, and whether you care only about heart rate or also about waveform morphology. A filter that is perfect for a consumer pulse-rate tracker can be too aggressive for pulse-wave analysis.

This article focuses on practical passband design for retaining useful PPG information while removing baseline drift and high-frequency contamination. It complements our articles on PPG Butterworth filter design, notch filter powerline noise in PPG, and PPG waveform basics.

Why PPG needs a bandpass stage

Raw PPG is dominated by a large DC component from static tissue and average blood volume, plus a much smaller AC pulsatile component tied to the cardiac cycle. On top of that, there may be respiratory modulation, vasomotion, ambient light leakage, electronic noise, and movement artifact.

A bandpass filter is attractive because it does two jobs at once:

• high-pass action reduces slow baseline drift
• low-pass action suppresses high-frequency noise

In many pipelines, it is the cleanest way to prepare the waveform for peak detection, heart-rate estimation, or more advanced denoising.

The physiologic bands that matter

A good passband begins with physiology.

Heart-rate band

For many adults, resting to exercise heart rates occupy roughly 0.8 to 3.5 Hz, though extreme cases can fall outside that range. Neonates, pediatric populations, and high-intensity athletes may require wider bounds.

Harmonics and morphology

The waveform is not a pure sinusoid. Sharp systolic upstroke and dicrotic structure introduce harmonics above the fundamental. If you care about morphology, a low-pass cutoff of 4 or 5 Hz may be too restrictive at some sample rates and sensor sites. Pulse contour work often benefits from wider low-pass limits, sometimes 8 to 12 Hz depending on the application.

Respiration and baseline components

Respiratory modulation often sits around 0.1 to 0.4 Hz. If your goal is only heart rate, a higher high-pass cutoff can remove more drift. If you also want respiratory information, the cutoff should be lower or the respiratory component should be estimated before aggressive filtering.

Common passband choices and what they imply

0.5 to 5 Hz

This is one of the most common general-purpose wearable settings. It usually works well for heart-rate estimation and basic peak detection. It removes slow drift aggressively enough to stabilize thresholds and rejects much of the high-frequency noise.

Tradeoff: it can attenuate very low heart rates and may soften morphology details.

0.3 to 8 Hz

A more morphology-friendly option. This preserves more low-frequency modulation and more harmonic content.

Tradeoff: more baseline residue and more high-frequency noise may pass through, especially in loose-wear wrist data.

0.7 to 4 Hz

Sometimes used in tightly scoped heart-rate monitoring pipelines where morphology is secondary and the motion environment is moderate.

Tradeoff: excellent robustness for rate extraction, but less suitable for pulse-shape studies.

The point is not that one band is correct. The point is that each passband encodes assumptions about the task.

Filter family selection

Butterworth

Butterworth filters are popular because of their smooth passband with no ripple. For PPG, that is a strong default because amplitude distortion is lower than with sharper ripple-prone alternatives. This is why Butterworth appears in so much wearable literature.

Chebyshev and elliptic

These can achieve steeper transitions with lower order, but the passband or stopband ripple can be undesirable for morphology-sensitive work. They may be acceptable in heart-rate-only systems where timing is more important than exact shape.

FIR versus IIR

Finite impulse response filters can deliver linear phase, which is appealing if waveform timing and shape preservation are critical. The downside is computational burden and longer filter lengths. Infinite impulse response designs are more efficient but often require careful handling of phase distortion.

In offline analysis, forward-backward filtering can remove phase distortion for IIR designs. In real-time wearables, that is not possible, so the phase characteristics of the chosen filter matter.

Order, transition width, and ringing

A steeper filter is not automatically better. Higher order brings sharper separation, but it also increases ringing risk, startup transient length, and sensitivity to implementation issues.

For most wearable PPG heart-rate pipelines, moderate-order designs are preferable. Engineers should pick the lowest order that achieves the needed attenuation. Overdesigned filters can look mathematically impressive while making real waveform interpretation worse.

Sampling rate and normalized design

Cutoffs should always be selected with the device sampling rate in mind. A 5 Hz low-pass behaves very differently at 25 Hz than it does at 250 Hz. Low-rate devices have less room between the cardiac band and Nyquist, so cutoff choices become more consequential.

This also affects morphology retention. At low sample rates, the waveform is already discretized coarsely. An overly conservative low-pass can erase the little shape detail that remains.

Zero-phase versus causal filtering

Offline research pipelines often use zero-phase filtering by applying the filter forward and backward. This is excellent for retrospective analyses because it removes phase distortion. However, it is not deployable in true streaming systems.

Real-time wearables need causal filters, which introduce delay and possibly phase distortion. If downstream algorithms depend on exact pulse-foot timing or beat-to-beat intervals, that delay needs to be understood and, if possible, compensated.

Interaction with peak detection

Peak detection performance depends heavily on bandpass choice. If the high-pass cutoff is too low, the residual baseline can destabilize adaptive thresholds. If it is too high, pulse amplitude may be reduced and low-heart-rate beats can flatten.

If the low-pass cutoff is too low, the systolic peak broadens and derivative-based detectors lose precision. If too high, high-frequency noise creates false candidates. This is why filter design and peak detection algorithms should be tuned together rather than independently.

Interaction with motion artifact

A bandpass filter helps with some motion artifact, but it is not a complete motion solution. When step cadence overlaps heart rate, frequency-based separation alone becomes weak. In these settings, the bandpass should be viewed as a front-end cleanup stage before adaptive filtering or Kalman filtering.

Benchmarking passband choices

When comparing candidate bandpass designs, evaluate:

• heart-rate MAE versus ECG
• beat detection sensitivity
• timing error of detected peaks
• waveform correlation on clean segments
• resilience during walking and running

Synthetic sweeps are useful, but real recordings are essential because wrist artifact often breaks ideal assumptions. A filter that looks great on clean bench data may underperform in free-living motion.

Practical recommendations by use case

Consumer heart-rate tracker

Start around 0.5 to 5 Hz with a moderate-order Butterworth design. Validate during rest, walking, and intermittent motion.

Morphology research

Use a wider passband, often something like 0.3 to 8 Hz or similar depending on sample rate, and validate dicrotic notch and derivative landmarks.

Respiration extraction from PPG

Preserve more low-frequency content or split the pipeline so respiratory modulation is estimated before aggressive cardiac-focused filtering.

Low-power embedded system

Prefer efficient IIR designs with stable coefficients and well-characterized delay. Avoid unnecessarily high order.

FAQ

What is the best bandpass filter for PPG?

There is no single best bandpass filter. For general heart-rate monitoring, a moderate-order Butterworth around 0.5 to 5 Hz is a common and practical starting point. Morphology applications usually need a wider band.

Why can a bandpass filter distort PPG morphology?

Because the PPG waveform contains useful harmonics beyond the heart-rate fundamental. If the low-pass cutoff is too low or the phase response is poorly handled, the filter can smooth peaks, alter notch structure, or shift timing.

Should I use FIR or IIR for PPG bandpass filtering?

Use FIR when linear phase is critical and computational cost is acceptable. Use IIR when efficiency matters, especially in wearable devices. Offline pipelines can use forward-backward IIR filtering to eliminate phase distortion.

Can a bandpass filter remove motion artifact from PPG?

It can reduce some motion contamination, especially when the artifact is outside the cardiac band. It is much less effective when motion overlaps heart rate, which is common during ambulation.

What cutoff frequencies should I use for wrist PPG?

A common starting point is 0.5 to 5 Hz for heart-rate work, but wrist devices vary widely in sampling rate, hardware, and fit. Final cutoffs should be validated on your own datasets.

Is zero-phase filtering okay for wearable development?

It is fine for offline research and benchmarking, but it is not a true real-time method. Streaming products need causal filters and should account for their latency and phase behavior.

References

• Allen J. Photoplethysmography and its application in clinical physiological measurement. DOI: https://doi.org/10.1088/0967-3334/28/3/R01
• Elgendi M. On the analysis of fingertip photoplethysmogram signals. DOI: https://doi.org/10.3390/rs6040156