Dr. Doyle’s Annotated PPG Bibliography

Allen separates the pulsatile AC signal from its slower DC baseline, where respiration and other low-frequency influences appear, and connects both to optics, instrumentation, and clinical use. This signal anatomy supplies vocabulary for the extraction papers that follow.

Castaneda treats wearable PPG as more than heart rate and pulse oximetry: waveform shape and its second derivative can inform vascular assessment, while wearable formats enable continuous monitoring. This is the applications-first bridge from sensor design to clinical use.

Tamura traces how wavelength-dependent absorption becomes SpO2, then examines circuits, calibration, motion, standards, and unobtrusive monitoring. The review grounds respiratory analysis in the pulse-oximetry instrument that produces many of the pleth waveforms researchers actually use.

Fine groups PPG error into patient, physiological, and external sources, including motion, ambient light, and contact pressure. The taxonomy is a practical guide for collecting metadata and designing stress tests before interpreting respiratory modulation as physiology.

Mejía-Mejía presents an end-to-end engineering workflow spanning acquisition, preprocessing, time- and frequency-domain analysis, machine learning, physiological estimation, and synthetic PPG generation. It connects raw-waveform storage to reproducible algorithm development and simulator-based testing.

Park organizes 118 studies around waveform formation, features and applications, noise, and signal processing. Its finding that preprocessing remains unstandardized makes the review a useful taxonomy for recording each transformation and comparing pipelines fairly.

Charlton follows wearable PPG from device design and waveform processing through clinical applications, research resources, and integration into care. The review serves as a systems checklist for moving a promising laboratory algorithm into daily-life monitoring.

Charlton's multi-author roadmap sets priorities across sensor design, signal processing, clinical applications, and enabling research. Its enduring value is the field-wide agenda for datasets, standards, bias, validation, and open science, rather than instruction in one algorithm.

Meredith links breathing to PPG through venous return, intrathoracic pressure, stroke-volume changes, and autonomic heart-rate effects. That physiological map explains why one waveform yields several respiratory surrogates and why their reliability varies across people and conditions.

Nilsson surveys respiratory rate, apnea and obstruction cues, and fluid-responsiveness modulation carried by the pleth, while emphasizing movement and vasomotor confounding. This is the clinical-breadth reference, not evidence for any single extraction technique.

Li compared six respiratory variations in 28 healthy participants across breathing rates and postures; the strongest surrogate changed with sex, posture, and rate. Preserve the raw waveform and evaluate several candidates instead of assuming one universal modulation.

Liu tested amplitude modulation, baseline wandering, frequency modulation, and filtering at six body sites during normal and deep breathing. Frequency modulation performed best overall, but measurement site and breathing pattern still altered respiratory energy and error.

Hartmann isolated measurement site in 36 healthy volunteers: frequency demodulation favored the forehead during normal breathing and the finger during deep breathing. Site and breathing pattern therefore belong in validation, not among interchangeable hardware choices.

Addison found DPOP increased with imposed breathing effort in seven healthy volunteers, suggesting modulation depth may complement respiratory rate. The cohort was small, and fluid level must remain stable during the analysis interval.

Addison evaluated 13 effort-sensitive features spanning amplitude, baseline, respiratory sinus arrhythmia, pulse transit time, and heart-rate shifts; six tracked imposed airway pressure across probes and loads. This is a feature-engineering map for breathing effort, not merely rate.

Nakajima built an earlobe monitor that adaptively separated cardiac and respiratory bands in real time, comparing outputs with ECG and transthoracic impedance at rest and during exercise. Maximum errors of 10 beats/min and 7 breaths/min mark a historical baseline.

Chon's VFCDM method tracked frequency modulation in 15 healthy participants breathing at 12–36 breaths/min and outperformed continuous-wavelet and autoregressive baselines. It remains an early time-frequency benchmark, not broad evidence of clinical performance.

Karlen's Smart Fusion extracts RIFV, RIIV, and RIAV, then rejects estimates compromised by artifact or disagreement among the three estimates. Evaluation in children and adults makes it the memorable prototype for quality-aware fusion in mobile pulse oximetry.

Charlton tested 314 extraction, estimation, and fusion combinations under ideal conditions and released the RRest toolbox and data. ECG was generally more precise than PPG, while time-domain estimation with modulation fusion ranked highly; artifact-rich validation remained necessary.

Pimentel combines multiple autoregressive orders across RIFV, RIAV, and RIIV, retaining estimates for over 90% of windows in two independent clinical datasets. Its durable lesson is robustness through modulation fusion and varied data, not one headline error.

Charlton maps the full pipeline: preprocessing, respiratory-signal extraction, rate estimation, fusion, quality assessment, and validation. Use this framework to locate each algorithm and decide whether its evidence supports wearable, clinical, or only controlled use.

Garde uses time-varying correntropy spectral density to estimate respiratory and heart rates from one spectrum, without demodulation or pulse-cycle detection. CapnoBase results expose the trade-off: 120-second windows improved error while reducing temporal resolution.

Park's adaptive lattice estimator uses heart rate to tune sequential IIR notches that remove the cardiac component and harmonics before tracking respiration. Its deliberately light architecture is the real-time counterpoint to more computationally intensive time-frequency methods.

Zhang models each PPG window as a sparse respiratory spectrum, tracks from the previous estimate, and uses an SQI to suppress unreliable windows. Operation at 10 Hz on CapnoBase makes it a concrete low-sampling-rate wearable design.

Dehkordi applies synchrosqueezing directly to PPG and reads instantaneous respiration from the dominant ridge between 0.1 and 1 Hz. CapnoBase's wide subject-level error range cautions against remembering only the low median RMS error.

Dehkordi uses synchrosqueezing first to recover RIIV, RIAV, and RIFV, then to track each instantaneous rate before peak-conditioned fusion. The fused estimate outperformed individual variations and simple fusion against capnography and nasal/oral airflow.

Cicone and Wu's deppG combines a de-shape short-time Fourier transform with synchrosqueezing to recover instantaneous heart and respiratory rates. CapnoBase and intense-exercise testing make it a rigorous motion-stress comparator, though not a lightweight embedded baseline.

Khreis scores amplitude, frequency, and baseline-wander modulations with respiratory quality indices, then fuses the strongest signals through a Kalman smoother. Tests in immobilized patients and daily activities anchor quality-aware tracking across contrasting motion conditions.

Iqbal combines selective windowing, conditioning, modified Welch filtering, and postprocessing for low-quality or interrupted PPG. The 90-second-window result is a trade-off, not a prescription: optimal window length and scaling can vary by subject and dataset.

Kim's adaptive IIR notch filter avoids overlapping moving windows and supports online respiratory-rate estimation. It is a lean real-time comparator to lattice filtering and Smart Fusion, although the evidence centers on simulation plus an implementation demonstration.

Pirhonen feeds four time-frequency representations of finger-PPG amplitude variability into a particle filter; wavelet synchrosqueezing performed best on VORTAL. Persistent artifact and non-respiratory components make this a Bayesian tracking benchmark with an explicit robustness limit.

Baker gives neural networks candidate respiratory rates plus respiratory quality indices; adding quality reduced mean absolute error by up to 38.17%. The model learns estimator reliability, not estimates alone, but still requires validation beyond the ICU.

Lei combines complementary ensemble empirical mode decomposition with ICA and non-negative matrix factorization to reconstruct cardiac and respiratory surrogates from PPG. On MIMIC/PhysioNet data, the framework outperformed digital filtering and conventional empirical mode decomposition.

Chowdhury's ConvMixer estimates respiratory rate directly from PPG with only 0.56 million parameters, offering a lightweight deployment example. Cross-dataset tests show that fine-tuning on a small target sample can improve out-of-distribution performance.

Hwang compares seven deep models, including a dilated residual network, across surgical-ICU PPG, BIDMC, and CapnoBase data. Uneven errors by dataset and breathing-rate group show how training-set balance and external validation govern apparent performance.

Shuzan offers a feature-engineered alternative to end-to-end networks: separately selected PPG features and Gaussian-process regressors estimate respiratory rate and SpO2. The targets required different best predictors, arguing against one shared feature set.

Chin compares 274 classical technique combinations with a CNN-LSTM using seven-second windows. The best classical pipeline slightly beat the neural model on BIDMC, while network transfer to CapnoBase makes the study a useful short-window benchmark.

Yang encodes PPG as sequential spikes in an end-to-end network with temporal feedback. On BIDMC, accuracy approached conventional deep models while estimated energy use was substantially lower, making this the low-power neuromorphic comparator.

Argüello-Prada introduces RISSDV, respiratory variation in successive systolic differences, alongside RIAV, RIIV, and RIFV. It lowered error and tolerated missed pulses in 53 stationary adults, but ambulatory and pediatric generalization remains untested.

Ponsiglione compares ECG- and PPG-derived respiration on iAMwell and CapnoBase. ECG R-peak features were markedly more accurate in these datasets, providing a useful counterweight to PPG-only optimism without establishing universal method superiority.

Jarchi uses co-located accelerometry to suppress motion in the Hilbert domain, reconstruct corrupted PPG, and estimate respiration with an autoregressive model. The architecture supports rest-to-motion monitoring instead of treating every motion-corrupted segment as unusable.

Lee combines 12 channels, multiple wavelengths, ICA, and truncated singular-value decomposition to reduce motion artifact during walking and running. Because validation targeted ECG-referenced heart rate, the findings guide hardware and artifact control, not respiratory-rate accuracy.

Elgendi compared eight signal-quality indices against expert labels; skewness best separated excellent, acceptable, and diagnostically unfit PPG in this dataset. It offers a simple, testable quality gate before respiratory estimation.

Sabeti derives six morphological features and uses a cost-sensitive SVM to classify clinician-annotated PPG quality without an ECG reference. Testing on 46 half-hour recordings from cardiopulmonary patients makes it a clinically grounded window-weighting strategy.

Orphanidou turns signal quality into a wearable control: its SQI identifies ECG and PPG segments suitable for reliable heart-rate estimation. SQI gating also reduced PPG respiratory-rate error and recording time with little loss of valid vital-sign data.

van Gastel learns a color-channel combination in the pulse band, then reuses it in the respiratory band to suppress motion in visible and infrared camera PPG. Adult guided-breathing and NICU tests establish a non-contact bridge, with dataset-specific results.

Zhang uses green PPG for heart rate and infrared PPG as an optical motion reference, followed by wavelet removal and reconstruction. The six-subject, 21-motion study informs multi-wavelength artifact design but validates heart-rate, not respiratory-rate, recovery.