World Sleep Industry ENCYCLOPEDIA

PPG sleep staging algorithms

4.4.1

Definition, optical measurement, and classification boundary

A PPG sleep-staging algorithm is software that uses a photoplethysmographic pulse waveform, with optional movement or other signals, to estimate wakefulness and sleep stages in successive time windows. PPG illuminates tissue and detects changes in transmitted or reflected light, primarily observing peripheral blood-volume changes associated with each heartbeat. It does not directly record electroencephalography, eye movements, or chin electromyography. Stage inference instead relies on statistical relationships between sleep and pulse intervals, waveform morphology, respiratory modulation, peripheral vascular tone, and activity. Its output is therefore an estimate based on cardiovascular and behavioral proxies, not a direct measurement of the electrophysiologic features used for conventional staging.

Outputs may be sleep versus wake, W/NREM/REM, W/light/deep/REM, or W/N1/N2/N3/R. Similar names do not guarantee equivalent definitions: “light” may combine N1 and N2, while “deep” often denotes N3 but may be proprietary. A system should disclose its classes, epoch duration, temporal alignment, contextual window, mapping to reference labels, and indeterminate-state policy. This entry addresses the PPG optical chain and its specific validation problems. The general framework belongs under sleep-staging algorithms, while radar-specific inference belongs to the next entry.

Optical hardware, wear location, and synchronized acquisition

PPG may use transmission or reflectance geometry at a finger, wrist, dorsal wrist, ear, or ring contact surface. Optical wavelength, emitted power, source-detector spacing, analog front end, dynamic range, sampling rate, automatic gain, and ambient-light rejection all affect the raw waveform. A watch, ring, and finger clip are not interchangeable merely because each supplies PPG. Development and validation records should identify the hardware generation, firmware, sensing site, orientation, fit, contact pressure, optical channels, original sampling, and any device-side resampling. Retaining only the final heart-rate series removes information needed to audit waveform quality and algorithm transfer.

When PSG supplies the reference, PPG, accelerometry, and PSG need a traceable common time base. Startup latency, clock drift, dropped packets, daylight-saving changes, or software-filled samples can align plausible waveforms with the wrong epoch. Synchronization, drift correction, interpolation, resampling, missing segments, and recording boundaries should be prespecified, with original timestamps retained for review. If the algorithm fuses acceleration, skin temperature, oxygen saturation, or another channel, the sampling, synchronization, missing-data policy, and contribution of each input should be stated separately. Multisensor performance must not be described as PPG-only performance.

Signal quality, features, and model pipeline

A typical pipeline includes dark-level or baseline handling, filtering, pulse detection, motion-artifact recognition, signal-quality scoring, and selection of valid segments. Models may use pulse intervals and their variability, pulse rate, amplitude, rise and decay morphology, pulse width, spectral or nonlinear features, and respiratory modulation, or learn temporal representations directly from raw or processed waveforms. A pulse interval derived from PPG is not synonymous with an electrical R–R interval. Ectopic beats, missed or duplicated peaks, and changes in pulse transit can enter HRV-like features. Corrections, windowing, context, normalization, and personal baselines should use only information available at the declared prediction time; future data or whole-night statistics must not leak into a claimed real-time output.

Quality gating is part of the algorithm, not a cosmetic post-processing step. Each epoch should distinguish valid, low-quality, missing, and algorithm-rejected data, and the system should state whether poor segments are removed, interpolated, assigned a coarser class, or handled by another sensor. Motion-based artifact removal must avoid erasing genuine autonomic or arousal responses, and a reconstructed “clean” waveform is not automatically a true observation. Developers should report attrition from recorded to analyzable nights, valid coverage per night, longest continuous gap, recovery after renewed contact, and whether gating selectively removes particular stages or participant groups.

Reference labels, data separation, and representation

Supervised labels should preferably come from synchronized PSG manually staged under an identified rule version, with scorer qualifications, blinding, consensus, and adjudication described. PPG carries stage information indirectly, and N1, brief wake, and stage boundaries are especially liable to confusion. When classes are merged, the mapping should be fixed before comparison rather than selected after results are seen. PSG artifacts or scorer disagreement should be marked as uncertain instead of treating a single label as error-free truth. Device-generated stages, existing clinical reports, or nonsimultaneous nights provide different levels of evidence and should not be presented as equivalent validation references.

Training, tuning, and final testing must be separated by participant. Multiple nights from one person, adjacent windows, and augmented versions of one waveform cannot cross partitions. Independence should also be examined by site, hardware generation, and acquisition period so the model does not learn a device or laboratory signature. Data should represent the intended ranges of age, sex, objectively characterized skin tone, body habitus, wear site, sleep efficiency, rhythm, medication, sleep disorder, and home or laboratory setting. Training only on healthy adults during quiet, single, signal-complete nights does not establish use in children, older adults, insomnia, severe sleep-disordered breathing, arrhythmia, or longitudinal home monitoring.

Performance, coverage, and night-level agreement

An evaluation should first state whether the input is PPG alone or PPG fused with acceleration or other signals, then use a synchronized PSG test set that was not used to train the model or tune thresholds. Epoch-level reporting should include a confusion matrix, per-stage recall, precision, F1, macro averages, chance-corrected agreement, and wake specificity, together with low-quality, missing, and rejected proportions. Accuracy calculated only among “valid epochs” can conceal substantial unusable data. Performance and coverage should therefore be presented together, with participant-level distributions and uncertainty intervals rather than only pooled epochs.

Night-level comparisons should address total sleep time, sleep-onset latency, wake after sleep onset, sleep efficiency, stage duration or proportion, REM latency, and transitions, using bias and limits of agreement to expose individual errors. High correlation does not mean numerical agreement, and high overall accuracy may merely reflect a high prevalence of sleep. Nights with low sleep efficiency, frequent awakening, different respiratory-event burdens, or atypical architecture require separate analysis. Evaluation should also determine whether staging error propagates into measures that use sleep time as a denominator. A system promoted for longitudinal trends additionally needs evidence on night-to-night repeatability, comparability after hardware replacement, and separation of real change from algorithmic variability.

PPG-specific failure modes and equity

Movement, wrist compression, loose fit, rotation, ambient light, detector saturation, sweat, hair, tattoos, cold temperature, poor perfusion, edema, and peripheral vascular conditions can alter the signal. Automatic gain may preserve a visible pulse while changing amplitude features. Atrial fibrillation, ectopic beats, paced rhythms, or missed beats can undermine features that assume regular intervals. The algorithm should distinguish these causes instead of labeling all failures as poor wear, and should be challenged with sustained compression, turning, cold exposure, rhythm irregularity, charging interruptions, and declining battery. When input requirements are not met, rejection, coarse fallback, or reacquisition guidance is preferable to silently completing a full hypnogram.

Melanin absorption, optical geometry, wavelength, and device control strategy may jointly affect the detected signal, but race or ethnicity is not a substitute for skin tone, perfusion, or sensing quality. Validation should use a clearly described objective skin-tone measure and record perfusion, temperature, site, contact, and hardware settings. Valid coverage, signal quality, rejection, and staging performance should be reported by relevant groups. When a disparity appears, investigators should locate it in acquisition, quality gating, reference labels, or the model rather than mask missing evidence by changing only a classification threshold. Equity analysis should also cover age, sex, body habitus, disease, and device fit, with underrepresented groups disclosed.

Use boundary, version governance, and acceptance

A consumer-stage chart can support personal trend observation within a validated range, but should not by itself diagnose or exclude disease, direct medication changes, or alter PAP or another treatment. Medical and research uses need an explicit intended purpose, reference standard, and human-review pathway. An algorithm validated to estimate stages has not thereby validated the same PPG stream for respiratory-event detection, oxygen measurement, or interpretation of cardiovascular abnormalities. Interfaces should expose missing signals, poor quality, rejection, class definitions, and known limitations rather than depict a probabilistic estimate as a precise brain state without uncertainty.

A change to the optical module, LED drive, automatic gain, wear mechanics, firmware, preprocessing, or model may invalidate earlier evidence. Every result should identify the hardware, firmware, model, rules, and application versions. Updates need bridging on frozen data, relevant subgroups, and the target device, followed by monitoring of coverage, stage distributions, night-level bias, complaints, and drift. Acceptance should obtain raw-input requirements, independent-test population, label mapping, per-stage results, invalid-data handling, privacy, and retention policies, then conduct synchronized, blinded local testing with representative users and devices. PPG waveforms and the heart rate, routines, and health-related features inferred from them are sensitive data whose purpose, access, retention, and secondary use should be limited.