Types of Noise in PPG Signals: Classification, Sources & Mitigation Strategies
Every PPG signal is a composite of the desired cardiac pulsatile component and multiple noise sources whose combined power typically exceeds the signal of interest by 10-100x at the raw sensor output. Understanding each noise type -- its physical origin, spectral characteristics, temporal behavior, and amplitude -- is essential for designing effective signal processing pipelines. The reason no single denoising algorithm works for all PPG applications is that different noise sources require fundamentally different mitigation strategies. A high-pass filter that removes baseline wander will not help with motion artifacts, and ambient light rejection circuitry does nothing for quantization noise.
This guide provides a systematic classification of PPG noise types with quantitative characterizations, physical explanations, and practical mitigation approaches. For a broader overview of PPG signal processing, see our PPG technology fundamentals and signal processing algorithms.
PPG Signal Composition and the Noise Floor
A raw PPG signal consists of several components superimposed. The DC component (static tissue absorption) represents 95-99% of the total signal and varies slowly with tissue characteristics, sensor placement, and ambient conditions. The AC component (cardiac pulsatile) represents only 0.5-5% of the total signal at the wrist and 1-10% at the fingertip, and contains the heart rate and waveform morphology information. Everything else is noise.
The perfusion index (PI), defined as the ratio of AC to DC component amplitude expressed as a percentage, is the simplest measure of signal quality. Typical PI values range from 0.02% in poorly perfused extremities to 20% at the fingertip under ideal conditions. At the wrist, where most wearable sensors operate, PI values of 0.1-2% are typical, meaning the cardiac signal is buried within a much larger baseline. Shelley et al. (2005) characterized PI distributions across measurement sites in 100 healthy subjects, finding median wrist PI of 0.8% compared to fingertip PI of 5.2% (DOI: 10.1213/01.ANE.0000153305.91152.BE).
This low PI at the wrist means that noise sources which would be negligible at the fingertip become significant at the wrist. The signal processing requirements for wrist-based wearables are fundamentally more demanding than for clip-on fingertip sensors, which is why the history of PPG wavelength selection has been driven largely by maximizing wrist-site AC amplitude.
Motion Artifacts
Physical Mechanisms
Motion artifacts are the dominant noise source in wearable PPG and arise from four distinct physical mechanisms that often act simultaneously:
Sensor-skin decoupling occurs when the device moves relative to the skin surface, changing the optical coupling between the LED/photodetector and the tissue. Even sub-millimeter displacement can alter the detected light intensity by 10-50% because the light collection geometry changes abruptly. This mechanism produces transient, high-amplitude artifacts that can saturate the photodetector or drive the signal below the noise floor.
Tissue deformation results from mechanical forces (compression, shear, stretching) applied to the tissue under the sensor. Pressure changes alter local blood volume by compressing superficial vessels, while stretching changes tissue optical path length. Teng and Zhang (2006) showed that 100 grams of applied force changes wrist PPG amplitude by 15-40% (DOI: 10.1088/0967-3334/27/12/003). This mechanism produces smooth, low-frequency artifacts correlated with the force waveform.
Venous blood redistribution occurs during arm motion when centripetal acceleration causes blood to pool or drain from superficial venous plexuses. Unlike the arterial pulsatile signal, venous volume changes are driven by gravity and inertial forces rather than cardiac contraction. During arm swing in walking, venous volume modulations at the wrist can produce optical signals comparable in amplitude to the arterial pulse, with frequencies that match the stride cadence.
Ambient light variation during motion occurs because body movement changes the angular relationship between the sensor and external light sources. Although shielded sensor designs minimize this, some ambient light leakage is unavoidable in wrist-worn form factors. This mechanism is discussed further in the ambient light noise section below.
Spectral Characteristics
Motion artifact spectral characteristics depend strongly on the type of activity. During walking at 3-5 km/h, the dominant motion frequency is 0.8-1.5 Hz (stride cadence) with harmonics extending to 4-6 Hz. During running at 8-12 km/h, the fundamental shifts to 1.3-1.8 Hz with harmonics to 8-10 Hz. During cycling, pedaling cadence of 60-100 RPM produces artifacts at 1.0-1.7 Hz. During weightlifting and resistance exercise, motion artifacts are typically aperiodic, broadband bursts during the concentric and eccentric phases.
The critical problem is spectral overlap with the cardiac signal. Resting heart rate occupies 0.8-1.7 Hz (48-100 BPM), and exercise heart rate extends to 3.3 Hz (200 BPM). During running, the stride cadence and its second harmonic often fall within 2-5 Hz, directly overlapping with exercise heart rates of 120-180 BPM. This spectral overlap is why simple bandpass filtering cannot separate cardiac from motion components. For detailed coverage of motion artifact removal algorithms, see our comprehensive motion artifact removal guide.
Quantitative Characterization
Benchmark studies quantify motion artifact severity using signal-to-motion-artifact ratio (SMAR). At rest, SMAR is effectively infinite (no motion artifacts). During slow walking, SMAR is typically 5-15 dB. During brisk walking, SMAR drops to 0-8 dB. During running, SMAR is commonly -5 to 5 dB (motion artifact power equals or exceeds cardiac signal). During vigorous arm exercises, SMAR can reach -15 to -20 dB.
The IEEE Signal Processing Cup 2015 dataset (Zhang et al., 2015, DOI: 10.1109/TBME.2014.2359372) provides standardized motion-corrupted PPG recordings from 12 subjects during treadmill running with concurrent accelerometer data, serving as the primary benchmark for motion artifact removal algorithm evaluation.
Baseline Wander and Low-Frequency Drift
Sources
Baseline wander refers to slow variations in the PPG DC component that occur over timescales of seconds to minutes. Multiple physiological and technical mechanisms contribute:
Respiratory modulation creates baseline oscillations at the breathing rate (0.15-0.4 Hz, approximately 9-24 breaths per minute). Respiration affects PPG through intrathoracic pressure changes that modulate venous return (respiratory-induced intensity variation, RIIV) and through sympathetic nervous system modulation of peripheral vasomotor tone (respiratory-induced frequency variation, RIFV). The respiratory component amplitude is typically 1-5% of the DC level. Notably, this "noise" can be a signal of interest -- respiratory rate extraction from PPG relies specifically on this modulation (Charlton et al., 2016, DOI: 10.1088/0967-3334/37/4/610).
Vasomotor activity produces spontaneous fluctuations in peripheral vascular tone mediated by the sympathetic nervous system. Mayer waves, oscillations in arterial blood pressure at approximately 0.1 Hz, produce corresponding PPG baseline variations with periods of 8-12 seconds. Traube-Hering-Mayer waves at approximately 0.03-0.04 Hz create even slower oscillations. These autonomic modulations can shift the PPG baseline by 5-20%.
Thermal drift in the sensor and tissue causes slow baseline changes over minutes to hours. As the LED heats the local tissue (typical LED thermal dissipation of 5-50 mW), local perfusion changes due to thermoregulatory vasodilation. The sensor's photodetector dark current and transimpedance amplifier offset also drift with temperature, though modern analog front-ends minimize this with integrated temperature compensation.
Contact pressure variation from wrist band loosening, arm repositioning, or gravitational effects when changing posture can shift the baseline by 10-50% over seconds. This is distinct from motion artifacts in that it occurs without rhythmic motion and persists until the contact condition stabilizes.
Mitigation
High-pass filtering is the standard baseline wander mitigation strategy. A cutoff frequency of 0.5 Hz removes most baseline wander while preserving heart rates above 30 BPM. However, this cutoff removes respiratory modulation, which is problematic if respiratory rate extraction is desired. Polynomial detrending (fitting and subtracting a low-order polynomial from each analysis window) provides an alternative that adapts to the specific baseline shape. Wavelet-based detrending using the approximation coefficients at decomposition levels 7-9 (for sampling rates of 100-250 Hz) offers better preservation of the low-frequency cardiac content (Raghuram et al., 2012, DOI: 10.1016/j.compbiomed.2012.09.005).
Ambient Light Interference
Sources and Characteristics
PPG photodetectors respond to all photons reaching the active area, not just those emitted by the sensor's LED. Ambient light sources introduce noise with characteristics that depend on the light type:
Sunlight produces a large, slowly varying DC offset and can saturate the photodetector entirely if the sensor lacks adequate optical shielding. Solar spectral irradiance at skin level can reach 100-1000 W/m^2 across the PPG-relevant wavelength range, overwhelming the 1-20 mW LED output by orders of magnitude. Indoor fluorescent and LED lighting typically produces 1-50 W/m^2 at the sensor surface.
Fluorescent lighting generates periodic intensity fluctuations at 100 or 120 Hz (twice the mains frequency due to the rectified waveform driving the phosphor). While these frequencies are well above the cardiac band and easily filtered, they can alias to lower frequencies if the PPG sampling rate is not sufficiently high. At a sampling rate of 100 Hz, a 120 Hz fluorescent artifact aliases to 20 Hz, which is still above the cardiac band. But at a 50 Hz sampling rate, a 60 Hz fluorescent artifact aliases to 10 Hz, closer to the region of interest.
LED-based artificial lighting driven by pulse-width modulated (PWM) drivers produces interference at the PWM frequency (typically 200 Hz - 20 kHz) and its harmonics. Some low-cost LED drivers operate at frequencies that can interact with the PPG sampling rate to produce low-frequency beat artifacts.
Hardware Mitigation
Modern PPG analog front-ends employ ambient light rejection through modulated illumination and synchronous detection. The LED is pulsed at a carrier frequency (typically 1-32 kHz), and the photodetector output is demodulated at the same frequency. This shifts the PPG signal away from the baseband ambient light spectrum, providing 40-60 dB of ambient light rejection. The Texas Instruments AFE4404 and Analog Devices ADPD4101, widely used in wearable PPG, implement this technique with integrated LED drivers and synchronous demodulators.
An alternative approach subtracts ambient-only measurements taken during LED-off intervals from LED-on measurements. This sample-and-subtract method provides approximately 20-30 dB rejection, limited by the temporal variation in ambient light between the on and off samples. Higher LED modulation frequencies improve rejection by reducing the time between on/off samples. For applications sensitive to ambient light, see our PPG conditions guide for clinical environment recommendations.
Powerline Interference
Characteristics
Electromagnetic interference from AC power lines couples into PPG systems at the mains frequency (50 Hz or 60 Hz, depending on region) and its harmonics. The coupling occurs through capacitive and inductive pathways between power wiring, the patient/subject, and the sensor electronics. The amplitude depends on proximity to power sources and grounding conditions, typically ranging from negligible to 5% of the DC component in clinical environments with multiple electrical devices.
For PPG, powerline interference is generally less problematic than for ECG because the PPG cardiac band (0.5-4 Hz) is well separated from 50/60 Hz. A simple notch filter at the mains frequency and its second harmonic (100/120 Hz) removes powerline interference without affecting the cardiac signal. Butterworth IIR notch filters of order 2-4 with quality factors of 30-50 provide 40-60 dB rejection with minimal spectral distortion at cardiac frequencies.
Practical Considerations
In battery-powered wearable devices operating far from power outlets, powerline interference is typically negligible. It becomes significant primarily in clinical settings where PPG sensors connect to mains-powered monitoring equipment, or when subjects are in contact with other mains-powered devices. Proper grounding, shielded cables, and differential signal routing minimize coupling. Software-based powerline removal using adaptive notch filters or spectral subtraction can track and remove interference even when the mains frequency drifts slightly from nominal.
Quantization Noise
ADC Resolution Effects
Quantization noise arises from the finite resolution of the analog-to-digital converter in the PPG signal chain. An N-bit ADC produces quantization noise with an RMS amplitude of LSB/(sqrt(12)), where LSB is the least significant bit voltage. The theoretical quantization SNR is approximately 6.02*N + 1.76 dB.
For PPG signals, the critical consideration is that the cardiac AC component is a small fraction of the total signal range. If the ADC input range is set to accommodate the full DC component, the AC component occupies only the bottom 1-5% of the ADC range. A 12-bit ADC (4096 levels) with the full range used for the DC component allocates only 40-200 levels to the AC component, yielding an effective AC signal resolution of approximately 5-7 bits and a quantization SNR of only 30-44 dB for the pulsatile signal.
Modern PPG analog front-ends address this by using high-resolution ADCs (20-24 bit) or by implementing DC subtraction (ambient light cancellation) before digitization. The Analog Devices ADPD4101, for example, provides programmable DC offset subtraction with a 14-bit ADC, effectively allocating the full ADC range to the AC component plus residual noise. This yields 14-bit effective resolution on the pulsatile signal, corresponding to quantization SNR exceeding 85 dB.
Oversampling and Noise Shaping
Oversampling by a factor of M improves effective resolution by log2(sqrt(M)) bits. Sampling at 1 kHz instead of 100 Hz (10x oversampling) and digitally decimating provides approximately 1.7 additional effective bits. Delta-sigma ADC architectures inherently combine oversampling with noise shaping, pushing quantization noise to higher frequencies where it can be removed by the decimation filter. Many modern PPG AFEs use delta-sigma ADCs for this reason, achieving 18-24 bit effective resolution at output data rates of 25-500 Hz.
Electronic Noise
Photodetector Noise
The photodetector (typically a silicon photodiode or phototransistor) contributes several intrinsic noise sources. Shot noise arises from the discrete nature of photon detection and photocurrent generation, with RMS current proportional to sqrt(2qI_DC*BW), where q is electron charge, I_DC is the mean photocurrent, and BW is measurement bandwidth. For typical PPG photocurrents of 0.1-10 microamps and bandwidths of 25-100 Hz, shot noise RMS is 0.3-6 nanoamps.
Dark current noise is shot noise from thermally generated carriers in the photodiode, producing a noise floor even without illumination. For silicon photodiodes at room temperature, dark current ranges from 0.01-10 nanoamps depending on active area and reverse bias. This sets the minimum detectable optical signal change and becomes significant for low-perfusion sites or very low LED drive currents.
Johnson-Nyquist (thermal) noise from the transimpedance amplifier feedback resistance contributes RMS voltage noise of sqrt(4kTRBW), where k is Boltzmann's constant, T is temperature, R is feedback resistance, and BW is bandwidth. For typical TIA feedback resistances of 100 kohm to 10 Mohm, thermal noise RMS is 0.5-50 microvolts.
The combination of these noise sources establishes the electronic noise floor of the PPG system. In well-designed analog front-ends, the electronic noise floor is 40-60 dB below the cardiac pulsatile signal, making electronic noise a minor contributor compared to motion artifacts and ambient light. However, for ultra-low-power systems with reduced LED drive current, electronic noise can become the limiting factor.
Signal Quality Assessment
Automated Quality Indices
Automated signal quality indices (SQI) classify PPG segments as usable or corrupted, enabling adaptive processing strategies that apply different algorithms based on signal condition. Several SQI approaches have been validated:
Template-matching SQI correlates each PPG beat with an average template derived from recent clean beats. Correlation coefficients above 0.9 indicate high quality, 0.7-0.9 indicate moderate quality suitable for heart rate extraction, and below 0.7 indicate corrupted segments. Elgendi (2016) validated template SQI on 104 patients with sensitivity of 93% and specificity of 97% for detecting usable segments (DOI: 10.1371/journal.pone.0168200).
Spectral SQI examines the ratio of spectral power in the cardiac band (0.5-4 Hz) to total spectral power. Clean PPG signals concentrate more than 80% of power in the cardiac band, while motion-corrupted signals have significant energy distributed across wider frequencies.
Perfusion-based SQI uses the perfusion index directly, with adaptive thresholds based on measurement site and individual baseline. PI values below 0.1% at the wrist typically indicate insufficient signal quality for reliable feature extraction.
Combining multiple SQI metrics through decision fusion (majority voting, weighted averaging, or machine learning classifiers) achieves the best performance. Orphanidou et al. (2015) demonstrated a multi-feature SQI achieving 95% accuracy in identifying segments suitable for heart rate estimation across diverse noise conditions on the MIMIC-II database (DOI: 10.1109/JBHI.2014.2338351). This quality assessment framework integrates naturally with downstream PPG processing algorithms to ensure reliable physiological parameter extraction.
Practical Noise Mitigation Pipeline
For engineers building PPG signal processing systems, the following pipeline addresses each noise type in the optimal order:
- Ambient light subtraction at the hardware level using synchronous detection or LED-off sample subtraction (40-60 dB rejection).
- DC offset removal and baseline wander correction using high-pass filtering at 0.3-0.5 Hz or adaptive polynomial detrending.
- Powerline interference removal using a notch filter at 50/60 Hz if the application environment requires it.
- Bandpass filtering to the cardiac band of interest (0.5-8 Hz for general PPG, 0.5-4 Hz for heart rate only).
- Motion artifact removal using accelerometer-referenced adaptive filtering or signal decomposition methods.
- Signal quality assessment to flag or reject segments where residual noise exceeds acceptable thresholds.
This ordering is important: each stage removes specific noise components to improve the conditions for subsequent stages. Applying motion artifact removal to a signal still containing baseline wander and ambient light interference degrades adaptive filter performance because the filter wastes degrees of freedom modeling non-motion noise sources.
For comprehensive guidance on implementing these processing stages, see our PPG signal processing algorithms reference and our motion artifact removal guide.