Definition, Uses, and Scope
A PPG sleep dataset organizes photoplethysmography acquired during sleep—often also called a blood-volume-pulse signal—with device, wear, timing, and quality metadata and reference labels such as sleep stages or respiratory events. It can support research on pulse and vascular rhythms, estimation of sleep and wake or sleep stages, assisted analysis of respiratory abnormalities, wearable evaluation, and reproducible algorithms. Reusable resources generally need continuous waveforms rather than only nightly summaries.
Raw PPG, device-calculated pulse rate, inter-beat intervals, heart-rate-variability features, and oxygen saturation are different data layers. SpO2 usually depends on multiwavelength optical signals, calibration, and device algorithms and cannot be treated as directly obtainable from any single-wavelength PPG file. Inter-beat intervals estimated from PPG are also not identical to ECG R-R intervals. A dataset should state what level is actually released and not call vendor sleep scores or a heart-rate sequence a raw PPG collection.
Optical Configuration, Wear Location, and Device Parameters
PPG may use transmission at the fingertip or reflection at the wrist, ring, or another site. Light-source wavelengths, LED currents, photodetectors, optical paths, contact area, and sampling schemes differ. Wear location changes blood-flow, motion, and pressure effects, so a laboratory finger clip and an everyday wristband are not the same distribution. Documentation should provide device and firmware, optical channels and wavelengths, sampling rate, ADC or quantization, gain, automatic exposure, analog and digital filtering, and timestamp precision.
Contact pressure, strap tightness, wrist orientation, skin temperature, perfusion, ambient light, hair, tattoos, and nail polish can affect signals or derived values. Sensor placement, wear instructions, reapplication, off-wrist periods, charging, disconnections, and sleep position should be recorded. If accelerometry, skin temperature, or electrodermal activity is also present, documentation should say whether it contributes to quality filtering or label generation. Processed waveforms without optical and wear configuration limit interpretation and reproduction across devices.
Data Levels, Synchronized References, and Labels
Data may include raw photodetector values, pulsatile components after DC removal, filtered BVP, pulse peaks, inter-beat intervals, heart rate, quality indices, SpO2, or window-level features. Publishers should define units, scale, missing codes, algorithm, and upstream source for every layer and retain raw or closest-to-raw waveforms. If only derived heart rate is available, users cannot assess peak-detection errors or redesign waveform-quality methods.
Sleep labels may come from technician-scored stages and respiratory events in synchronized PSG, automated device results, diaries, or inference from another sensor, with different evidentiary strength. Alignment should document a common time source, start delay, clock drift, resampling, and missing intervals. The PPG pulse follows ECG electrical activity after a physiological transit delay, and oximeters may add internal smoothing delay. Reference signals must not be aligned merely by matching the first rows of files.
Signal Quality, Preprocessing, and Version Management
Common problems include motion artifact, low perfusion, ambient-light leakage, pressure changes, clipping, baseline drift, missed or double pulse detections, off-wrist periods, and long gaps. Quality flags should distinguish not worn, weak signal, gross motion, hardware failure, and algorithmic rejection rather than treating zero-filled or smoothed sequences as valid physiology. Algorithm evaluation should report usable coverage and rejection rate so that performance is not presented only for the cleanest intervals.
Preprocessing can include detrending, band-pass filtering, motion correction, resampling, normalization, peak detection, and repair of abnormal inter-beat intervals. Every step needs parameters, code, and order. A manifest should connect participant, night, device, raw signals, synchronized labels, and quality data and include a dictionary, checksums, license, change log, and known issues. Firmware and vendor-algorithm updates can alter derived values and should create traceable releases instead of silent replacement.
Population Representation, Measurement Bias, and Generalization
Optical performance can vary with skin pigmentation, skin thickness, temperature, peripheral circulation, smoking status, motion, and wear conditions. Datasets should use appropriate methods to describe relevant physiological and device conditions where feasible and examine quality and error by skin tone, age, sex, condition, perfusion, and device. Race labels are not precise substitutes for skin tone, and an overall average must not hide systematic error in a subgroup.
Laboratory PSG samples often come from referred patients, whereas consumer-device data can favor people able and willing to wear and share data over long periods. One hospital, age group, model, or wrist location limits external applicability. Many windows do not equal many independent participants, and a few individuals may contribute many nights. Publishers should report people, nights, devices, quality, condition spectrum, and label distributions together and retain realistic motion and failure scenarios.
Machine-learning Splits, Evaluation, and Reproducibility
Training, validation, and testing must be partitioned by participant, keeping a person's repeat nights, overlapping windows, alternate features from one waveform, and augmented copies together. Random window splitting leaks individual pulse morphology, device response, and nightly context. Filtering, quality thresholds, peak detection, normalization, imputation, and model thresholds must use training information only. Tests should remain independent and add cross-device, cross-site, and cross-population validation.
Metrics should fit the task. Pulse detection requires event matching and inter-beat-interval error; sleep staging requires class-specific performance, confusion, and agreement; respiratory events or SpO2 estimation require comparison with the appropriate reference. In addition to mean performance, reports should include quality coverage, rejection, low-perfusion and motion performance, and uncertainty intervals. Reproduction requires data and label releases, partition manifests, preprocessing, dependencies, random seeds, and final thresholds without tuning after repeated test inspection.
Privacy, Access Governance, and Clinical Limits
Continuous PPG can reveal cardiac rhythms, routines, activity, and possible health status and can be linked with time, devices, location, or other sensors. Publishers should perform data minimization, de-identification, and re-identification risk assessment consistent with consent and applicable rules and choose open, application-based, or controlled access according to sensitivity. Users must protect waveforms, credentials, and derived models; comply with licensing and restrictions on re-identification or redistribution; and address withdrawal, correction, retention, and access expiry.
PPG algorithms can support research and assisted monitoring but are not automatically equivalent to PSG, a clinical pulse oximeter, or diagnosis. A sleep stage inferred from PPG is a model estimate, not a direct observation of an EEG stage, and abnormal heart-rate, oxygen, or breathing outputs require the device's intended scope and other clinical information. Medical use also needs target-population and hardware-version validation, quality-failure alerts, appropriate professional review, and applicable regulatory processes. Mature datasets make these evidentiary limits explicit.