📄 / / profiles / r-and-d / raw_evidence / scientific-agent-skills / skills / neurokit2 / references / ppg.md

Photoplethysmography

Photoplethysmography

Checked 2026-07-23 against NeuroKit2 0.2.13 stable runtime/source, the official PPG API, 0.2.13 release notes, and measurement guidance.

Acquisition contract

Record sensor mode (reflectance/transmission), wavelength(s), anatomical site, attachment/contact pressure, device/firmware, raw unit/range, ambient-light handling, sampling rate, clock, temperature/perfusion, activity/posture, and motion/accelerometer channels. Validate across representative skin pigmentation, anatomy, age, vascular state, motion, and intended population.

PPG is optical blood-volume-pulse measurement, not cardiac electrical activity. Pulse timing and morphology depend on site and vascular/transit dynamics.

Stable high-level pipeline

signals, info = nk.ppg_process(
    ppg,
    sampling_rate=100,
    method="elgendi",
    method_quality="templatematch",
)

Pinned 0.2.13 default columns:

PPG_Raw, PPG_Clean, PPG_Rate, PPG_Quality, PPG_Peaks

info contained PPG_Peaks, sampling_rate, and peak/correction method metadata. The live doc's older codebook can omit PPG_Quality; stable runtime/source includes it.

Cleaning and peak methods

For explicit selection:

clean = nk.ppg_clean(
    ppg,
    sampling_rate=100,
    method="elgendi",
)
markers, peak_info = nk.ppg_peaks(
    clean,
    sampling_rate=100,
    method="elgendi",
    correct_artifacts=False,
)

Stable cleaning methods include:

elgendi;
nabian2018 (can use expected heart rate);
langevin2021;
goda2024; and
none.

Stable peak methods include:

elgendi;
bishop (peaks plus pulse onsets);
charlton (MSPTDfast v2; peaks plus onsets); and
charlton2024 (superseded v1).

Method-dependent extra outputs are a primary reason not to hard-code one schema. Validate peak/onset performance against labeled data at the actual site, rate, perfusion, motion, and population.

correct_artifacts=True uses the cardiac peak-correction path. Retain raw/corrected peaks and correction categories; correction cannot repair a low-quality optical waveform.

Quality outputs added/expanded in 0.2.13

quality = nk.ppg_quality(
    clean,
    peaks=peak_info["PPG_Peaks"],
    sampling_rate=100,
    method="templatematch",
)

Stable quality methods and scales differ:

templatematch: continuous similarity, typically 0–1;
dissimilarity: unbounded, where zero is highest similarity;
ho2025/interval-consistency path: binary interval quality;
skewness, kurtosis, entropy: unbounded windowed metrics;
perfusion: percentage-like 0–100 and requires raw PPG; and
relative_power: 0–1, requires raw PPG, and defaults to 60 s windows.

No threshold is universal across these outputs. Name the method and direction/scale. Short signals can be invalid for a method's default window. ppg_process() passes peak indices—not the whole info dict—to quality estimation.

Quality relative to an average pulse does not prove physiological accuracy: repeated motion-corrupted beats can be morphologically consistent. Combine morphology, motion, contact/perfusion, missingness, detector agreement, and endpoint-specific validation.

Sampling and preprocessing

Sampling requirements depend on endpoint:

pulse rate needs less bandwidth than morphology, onset timing, or derivatives;
wrist wearables commonly use lower rates than laboratory finger systems;
resampling cannot recover missing onset precision or a dicrotic feature; and
a nominal rate does not establish clock accuracy or anti-alias filtering.

Do not prescribe one universal minimum. Validate rate and filters for the exact detector/feature and report native/processed rates. Process at native rate before multimodal alignment when possible.

Motion, contact pressure, ambient light, vasoconstriction, temperature, pigmentation, site, and clipping can all change amplitude/morphology. Preserve a quality/artifact mask and accelerometry where available. Avoid interpolating through corrupted pulses.

PPG-derived variability is PRV

prv = nk.hrv_time(peak_info, sampling_rate=100)

NeuroKit2 accepts PPG peaks in HRV functions, but interpretation remains pulse-rate variability. PRV contains pre-ejection and pulse-transit variability and can differ from ECG HRV by site, posture, respiration, activity, temperature, and vascular state.

For an HRV-equivalence claim:

acquire synchronized ECG and PPG;
validate pulse/beat matching and lag/drift;
prespecify agreement metrics and acceptable error for each HRV endpoint;
test rest, task/activity, motion, and relevant populations/sites; and
report PRV terminology when equivalence is not established.

Do not infer ECG morphology, rhythm diagnosis, oxygen saturation, blood pressure, or arterial stiffness from this basic PPG pipeline.

Event and interval analysis

epochs = nk.epochs_create(
    signals,
    events,
    sampling_rate=100,
    epochs_start=-1,
    epochs_end=10,
    baseline_correction=False,
)
event_features = nk.ppg_eventrelated(epochs)
interval_features = nk.ppg_intervalrelated(signals)

Documented event fields include baseline/min/max/mean/SD rate, times, and polynomial trend coefficients. Interval output includes mean rate and HRV-family columns. Availability depends on input columns, duration, and release; inspect runtime output.

Pulse morphology

ppg_segment() returns a dict of pulse epochs. Morphology comparisons require:

consistent site, attachment, pressure, wavelength, and polarity;
validated onsets/peaks and quality masks;
appropriate baseline/amplitude normalization;
sufficient sampling/bandwidth;
control of heart rate and vascular state; and
endpoint-specific evidence.

A dicrotic feature in a processed waveform does not by itself validate aortic valve timing or arterial stiffness.

Interpretation boundary

Use these tools for research and education. They are not validated here for arrhythmia, oxygen saturation, blood pressure, disease detection, remote patient monitoring, alarms, or wearable medical-device validation.