PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Dataset Transfer and Reliability of Explainable Artificial Intelligence for RhythmFormer Remote Photoplethysmography

Louis Chen, Torbjörn E. M. Nordling

arXiv:2609.03663v1cs.CVcs.AIeess.IV

TL;DR

The paper addresses the limited quantitative evidence behind visual rPPG explanations and asks whether attribution methods transfer across datasets and track model reliability. It evaluates four explanations across eight condition-specific NCKU-rPPG models and UBFC-rPPG using skin coverage and SaCo. The results show that Beyond Intuition ranks highest in the reported cross-dataset comparison, while attribution measures are complementary to performance rather than proxies for it.

  • Problem

    Visual plausibility alone does not establish that highlighted facial regions affect an rPPG prediction or that attribution ordering agrees with spatial perturbation effects.

  • Method

    The study evaluates raw attention, rollout, attention flow, and Beyond Intuition across eight NCKU-rPPG scenarios and UBFC-rPPG using skin coverage and SaCo.

  • Results

    Beyond Intuition ranked highest on the reported datasets, while most within-participant attribution–performance coefficients were below |ρ| = 0.10 and only its coverage followed performance across scenarios.

  • Takeaways & Limitations

    Skin coverage and SaCo carry information complementary to performance measures, so attributing to skin does not by itself guarantee an accurate estimate.

  • Takeaways & Limitations

    The training protocol is a deliberate limitation, and the analyses do not test cross-dataset generalisation.

Abstract

from arXiv · show

Background. Remote photoplethysmography estimates the cardiovascular pulse from facial video, and its explanations have rested on inspecting heatmaps rather than on quantitative evidence about where a model reads it. We quantified the explanations and asked whether such explanations transfer between datasets and track model performance. Method. We trained eight condition-specific RhythmFormer models on NCKU-rPPG, recorded under three illumination levels, speaking, rotation, and cycling, estimated one heart rate per 5.12-second clip, and set them beside a UBFC-rPPG reproduction. Raw attention, rollout, attention flow, and Beyond Intuition were assessed by skin coverage and the Salience-guided Faithfulness Coefficient (SaCo). Results. Beyond Intuition ranked highest on both datasets, at median coverage 0.789 and SaCo 0.837 on Static level 3 against 0.826 and 0.917 on UBFC-rPPG; lower ranks differed. Within one participant of one condition, neither measure was related to a clip's heart-rate error, waveform correlation, or signal-to-noise ratio on either dataset: 186 of the 252 coefficients fell below $|ρ|=0.10$ and 28 reached $p<0.05$ against the 13 expected by chance. Across the eight scenarios only Beyond Intuition's coverage followed the three performance measures, at $ρ=-0.43$, $+0.57$, and $+0.43$, while the attention-only methods' SaCo ran opposite to each. It failed at 40 lux alone, its median coverage falling to 0.180 and its median SaCo to $-0.178$, whereas motion degraded the estimates far more without such a drop. Conclusions. Skin coverage and SaCo carry information complementary to the performance measures rather than a proxy for them: attributing to the skin does not guarantee an accurate estimate. What an attribution reveals about a condition is where the model looks rather than how faithfully its map is ordered.

1 Introduction

The study asks whether quantitative XAI explanations for rPPG reveal model behaviour, transfer across datasets, and track reliability. It evaluates these questions across controlled NCKU-rPPG scenarios and UBFC-rPPG.

  • Motivation: rPPG recovers cardiac pulse from ordinary facial video, but weak pulse-induced skin-colour changes can be confounded by illumination, camera characteristics, and motion.These factors complicate reliable interpretation of facial-video predictions.
  • Motivation: Dataset differences in recording characteristics, physiological references, sampling rates, and participant composition can prevent models trained on one dataset from generalising to another.The paper treats dataset adaptation as part of model evaluation.
  • Study aims: The study investigates whether XAI explanations are faithful to rPPG model behaviour and whether they transfer across datasets and track model reliability.It focuses on where models obtain physiological evidence and whether attribution rankings recur.
  • Study design: The authors compare XAI methods on Static level 3 and UBFC-rPPG, then relate median SaCo across eight NCKU-rPPG scenarios to error, waveform correlation, and signal-to-noise ratio.The design uses participant-independent splits and eight independently trained checkpoints.
  • Scope: The study supports descriptive evidence about transfer and reliability response rather than significance, causality, or cross-dataset generalisation claims.This scope limitation constrains how the comparisons should be interpreted.

2 Related Work

Prior rPPG XAI work has mainly used visualisations whose plausibility does not establish causal or perturbation-based faithfulness. This study extends a quantitative skin-coverage and SaCo framework from one controlled dataset to independently trained models across eight scenarios.

  • Prior XAI work: Existing rPPG XAI studies commonly visualise attention, gradient saliency, Grad-CAM, or channel–spatial interaction maps.These approaches have been applied to models including Dual-path TokenLearner, PhysKANNet, DD-rPPGNet, TS-CAN+, and CIN-rPPG.
  • Evaluation gap: Visual plausibility alone does not establish that highlighted regions affect an rPPG prediction or that attribution ordering agrees with spatial perturbation effects.The paper identifies quantitative perturbation faithfulness as an unresolved evaluation need.
  • Prior quantitative framework: A prior UBFC-rPPG evaluation paired skin coverage with an rPPG-adapted Salience-guided Faithfulness Coefficient.It examined whether attention aggregation restores multi-hop paths excluded by sparse top-k routing.
  • Prior quantitative framework: Beyond Intuition previously achieved the strongest median skin alignment and SaCo result, but only on one dataset, one model family, and a limited recording setting.That result motivates replication rather than a general XAI ranking.
  • Current study: The present study extends the framework to eight scenario-specific RhythmFormer models covering illumination, speaking, head rotation, and cycling.It tests transfer across independently trained models and examines whether multi-hop leakage recurs under varied conditions.

3 Methods

The methods use NCKU-rPPG recordings paired with synchronized physiological signals, controlled illumination and motion scenarios, and a conversion pipeline that produces facial videos for RhythmFormer analysis.

  • Acquisition: The acquisition system recorded frontal video at 1920×1080 pixels and 50 frames per second alongside ECG at 500 Hz and finger PPG at 125 Hz.Alignment files paired selected video frames with PPG measurements at the source video cadence.
  • Experiment design: The experiment analysed static illumination levels 1, 3, and 5, speaking and rotation at level 3, and cycling at all three illumination levels.Cycling varied cadence and resistance under researcher supervision.
  • Experiment design: Illumination targets were 40±50, 200±50, and 700 ± 50 lux for levels 1, 3, and 5, respectively.Static recordings limited voluntary movement, while speaking, rotation, and cycling introduced distinct motion demands.

3.2 Face extraction and resizing

Face extraction converts aligned source videos into standardized 128×128 facial crops for RhythmFormer training and evaluation. The pipeline detects faces, estimates landmarks, smooths crop geometry, validates encoding, and preserves aligned sequences.

  • Sequence preparation: The pipeline reads each Camera 1 video and alignment CSV, retains the shorter valid video–PPG sequence, and preserves the 50 Hz cadence rather than resampling to 30 Hz.The retained sequence begins at the aligned start frame and uses contiguous finite PPG values.
  • Face localisation: MTCNN detects faces frame by frame, and FaceXFormer estimates 68 landmarks used to define a square rPPG crop.The crop is shifted upward by eight percent of its side to include more forehead.
  • Resizing: The processed square crop is shifted within the frame, extracted, and resized from approximately 510 × 640 pixels to 128 × 128 pixels.The crop geometry is recomputed from each frame’s landmark span.
  • Crop tracking: Crop centres and side lengths are interpolated across missing detections and smoothed with median and Savitzky–Golay filters.Each session required a raw face-detection rate of at least 0.80.
  • Validation: Lossless FFV1 outputs were accepted only when encode–decode temporal RGB correlation reached 0.99, which every one of the 607 accepted sessions met.The converter deleted outputs and stopped for sessions below the threshold.

3.3 Participant split and RhythmFormer preprocessing

The study split participants before scenario selection, preventing participant leakage, and applied fixed preprocessing to converted facial-video and PPG sessions.

  • Participant split: One participant was excluded because the required alignment CSV files were absent, leaving 77 participants in the converted dataset.
  • Participant split: 65 participants and 511 sessions were assigned to training, while 12 participants and 96 sessions were assigned to testing.
  • Participant split: Scenario filtering followed the participant split, so different sessions from one participant could not enter both partitions.
  • Participant split: Each scenario retained 63–65 training participants and all 12 test participants after missing-session filtering.
  • Preprocessing: The pipeline used already cropped 128×128 FFV1 videos, converted channel order, and performed no additional face detection, tracking, or cropping.
  • Preprocessing: Sessions were standardized, divided into non-overlapping 256-frame clips, and processed without temporal resampling.At 50 Hz, each clip lasted 5.12 s versus approximately 5.33 s for the UBFC-rPPG format.

3.4 RhythmFormer architecture

RhythmFormer converts RGB facial video into a one-dimensional rPPG waveform through a hierarchical temporal architecture designed to expose periodic structure across scales.

  • Architecture: The Fusion Stem combines raw frames with temporal differences, while a three-dimensional convolution embeds non-overlapping spatial patches.
  • Architecture: Three hierarchical Temporal Periodic Transformer stages each contain two Periodic Sparse Attention blocks.
  • Architecture: Stage-specific temporal downsampling and upsampling expose periodic structure at multiple temporal scales.
  • Output: After the six attention blocks, spatial averaging and a one-dimensional predictor produce one waveform value per input frame.
  • Periodic Sparse Attention: Periodic Sparse Attention routes selected keys and values using coarse scores before refined attention, alongside a parallel local-context branch.

3.5 Scenario-specific training

The experiment trained independent scenario-specific RhythmFormer models with a shared configuration, composite objective, and heart-rate-aware temporal augmentation.

  • Training design: Eight independent RhythmFormer models were trained, each using only training clips from its scenario and testing on corresponding held-out scenario clips.
  • Training objective: The training objective combined negative-Pearson waveform loss, frequency-domain heart-rate cross-entropy, and heart-rate-distribution Kullback–Leibler loss.
  • Augmentation: Heart-rate-aware temporal augmentation was independently applied to each clip with probability 0.5.
  • Augmentation: Low-heart-rate clips were accelerated, high-heart-rate clips were temporally expanded, and clips from 75 to 90 beats per minute remained unchanged.
  • Augmentation safeguards: Augmented labels with non-finite values or standard deviation no greater than 10^-8 were restored to their original video–label pairs.
  • Model selection: Formal testing used the final training epoch because no validation loader or validation-based model selection was performed.

3.6 Performance evaluation

Performance was evaluated per short clip using standardized waveform processing, multiple heart-rate estimators, and participant-weighted error, correlation, and signal-to-noise metrics.

  • Analysis unit: Each 256-frame clip yielded one heart-rate estimate, preserving approximately 5-second temporal resolution rather than averaging clips into recording-level values.The NCKU-rPPG clips lasted 5.12 s, while the UBFC-rPPG reproduction used approximately 5.33-second clips.
  • Heart-rate estimation: Reference heart rate was estimated from each PPG clip using the dominant spectral peak and the median interval between detected systolic peaks.
  • Heart-rate estimation: 0.65 to 1.09 beats per minute was the median absolute difference between the two reference estimators across stationary conditions.
  • Heart-rate estimation: Cycling produced median estimator differences of 1.81–2.47 beats per minute and mean differences of 5.49–8.54 beats per minute.30 of 2,559 cycling clips reached the 150-beats-per-minute search-band boundary, whereas no stationary clip did.
  • Metrics: Predicted and reference waveforms were detrended, bandpass filtered from 0.75 to 2.5 Hz, and evaluated with MAPE, Pearson correlation, and SNR.
  • Metrics: All estimates used participant-weighted means, averaging clips within participants before averaging participant means.
  • Correlation: Waveform correlation was computed within clips at a fitted synchronization lag, distinguishing it from heart-rate correlation across clips.The fitted lag corrected constant recording-level synchronization offsets and left spectral magnitude-based metrics unchanged.
  • SNR: 75.2% of NCKU-rPPG clips had reference heart rates above 75 beats per minute, making the SNR measure closer to a first-harmonic measure than its formula suggests.

3.7 Explainable artificial intelligence attribution methods

The paper compares four attribution methods for RhythmFormer, differing in whether they propagate attention across layers, retain temporal structure, and incorporate gradient-based feedback.

  • Raw attention: Raw attention measures layer-local source-position use without propagating attribution across layers.Received attention sums the head-averaged attention matrix over query positions before spatial aggregation.
  • Attention rollout: Attention rollout augments attention matrices with residual connections and multiplies them in forward order to estimate cross-layer propagation.The residual weight is 1/2, and rollout operates on temporally collapsed spatial matrices.
  • Attention flow: Attention flow models stacked attention and residual connections as a layered capacity graph, assigning each source position the maximum flow reaching output positions.Its spatial nodes are the same temporally collapsed 4 × 4 or 8 × 8 representations used by the other attention-based methods.
  • Beyond Intuition: Beyond Intuition combines projection-corrected rollout with an integrated-gradient reasoning-feedback mask.Unlike the other methods, it retains temporal coordinates, unifies refined-attention blocks at T = 20, and pools time afterward.
  • Implementation: The rotation scenario required FP32 attribution because mixed-precision attribution produced non-finite values.This change affected attribution generation only, not the checkpoint, parameters, or training procedure.

3.8 Anatomical plausibility and perturbation faithfulness

Explanations were evaluated with skin coverage for anatomical plausibility and SaCo for perturbation faithfulness, using method-specific spatial resolutions and clip-level aggregation procedures.

  • Skin coverage: Skin coverage measures the proportion of normalized attribution assigned to pixels in a binary face-skin and nose mask.The mask came from a pretrained BiSeNet face parser, and coverage was explicitly not treated as faithfulness.
  • SaCo: SaCo measures perturbation faithfulness by ranking spatial positions by salience and dividing them into K = 8 equally sized groups.Each group is masked independently in every video frame, and waveform perturbation impact is measured using mean absolute error.
  • Map resolution: Beyond Intuition supplied an 8 × 8 attribution map to SaCo, while the three attention-only methods supplied 4 × 4 source maps.All maps were min–max normalized before bicubic upsampling to 128 × 128 pixels.
  • Aggregation: Attribution visualizations averaged clips within each scenario, whereas violin and scatter plots retained clip-level observations.Skin-coverage violins averaged each clip across TPT1–TPT3; scatter plots used refined TPT3 coverage for attention-only methods and one value for Beyond Intuition.
  • Visualization: Combined visualizations separated UBFC-rPPG from the eight NCKU-rPPG scenarios and encoded dataset or scenario without pooling scenario densities.Scatter plots used circles for NCKU-rPPG clips and crosses for UBFC-rPPG clips.

3.9 Cross-dataset transfer and scenario-level reliability comparisons

The study compares attribution summaries with performance within participants and across scenarios, preserving the nested data structure and treating scenario-level associations descriptively.

  • Scope and limitations: The two evaluations were not treated as a controlled cross-dataset experiment because they retained different participants, cameras, preprocessing, frame rates, and checkpoints.UBFC-rPPG SNR was recomputed from archived waveforms using the same estimation procedure as the NCKU-rPPG evidence.
  • Comparison design: Three clip-level performance measures—heart-rate MAE, waveform correlation, and SNR—were related to skin coverage and SaCo, producing six relationships plus one between the attribution summaries.The analyses covered eight NCKU-rPPG scenarios and UBFC-rPPG.
  • Within-participant analysis: Within-participant correlations were averaged over participants after computing each participant’s Spearman coefficient.This avoided measuring primarily between-participant differences in heart rate, skin tone, and facial geometry.
  • Pooling caveat: Pooling clips produced a Static level 3 waveform Pearson–skin coverage correlation of −0.41 with p < 10^-12, whereas within-participant coefficients ranged from −0.16 to +0.13 and were unresolved.The pooled relationship therefore differed sharply from the clip-level within-participant analysis.
  • Scenario-level analysis: Scenario-level analysis related each method’s median SaCo and skin coverage across eight NCKU-rPPG scenarios to MAE, waveform correlation, and SNR.These eight independently trained conditions were summarized descriptively without significance tests or causal interpretation.

4 Results

Performance varied strongly across illumination and motion conditions, while attribution rankings and attribution–performance relationships were often unstable across methods and datasets.

  • Performance across conditions: Rotation degraded performance more than speaking, with heart-rate error about 1.4 times as large and SNR 6.3 dB lower.Both motion conditions performed worse than the static level 3 model trained and tested at the same nominal illumination.
  • Performance across conditions: The three cycling conditions were weakest on every performance column, with mean absolute errors of 15 to 19 beats per minute and mean SNR from −8.9 to −11.9 dB.Bike level 5 was nominally best among cycling conditions, but the worst condition depended on the performance measure.
  • Performance relationships: Across scenarios, MAE ranked against heart-rate correlation at ρ = −0.98, waveform correlation at ρ = −0.95, and SNR at ρ = −0.88.Waveform correlation and SNR ranked together at ρ = +0.98, but they were not interchangeable.
  • Attribution summaries: Across all nine conditions, refinement raised raw-attention coverage and lowered attention-flow coverage without exception, while lowering attention-rollout coverage in eight.Static levels 1 and 3 were unusual because these directional changes were too small to resolve.
  • Attribution–performance relationships: Beyond Intuition accounted for 14 of the 25 directional attribution–performance pairs and all five SNR pairs.Where relationships resolved, attributed skin could accompany larger error, while SNR–skin coverage was negative in the two resolved cases.

5 Discussion

The discussion finds that attribution measures provide information complementary to heart-rate reliability rather than serving as proxies for it. Beyond Intuition generally ranked highest, but its skin-focused attribution failed in dim light, showing that where a model looks does not guarantee accurate estimates.

  • Attribution and reliability: 186 of 252 clip-level coefficients had |ρ| below 0.10, while 28 reached p < 0.05 versus 13 expected by chance.Neither attribution measure was generally related to heart-rate error, waveform correlation, or signal-to-noise ratio within a participant and condition.
  • Method comparison: Beyond Intuition supplied 14 of 25 directional relationships, exceeding the three attention-only methods combined.It also had the highest median refined skin coverage in seven of nine conditions.
  • Across-scenario patterns: Across eight scenarios, only Beyond Intuition’s skin coverage followed reliability measures, with ρ = −0.43 for error and +0.57 and +0.43 for waveform correlation and SNR.Attention-only methods’ SaCo showed the opposite direction for each reliability measure.
  • Method comparison: Attribution methods are not interchangeable: raw attention and rollout agreed on SaCo ordering at ρ = 0.97 but on attribution location at only 0.18.Rollout and attention flow showed the reverse pattern, agreeing more in coverage than in SaCo.
  • Condition effects: Speaking, rotation, and cycling raised heart-rate error from 1.96 ± 0.34 to 8.9 ± 2.6–16.1 ± 2.3 beats per minute while Beyond Intuition’s attribution remained comparatively stable.These motion conditions also produced lower waveform correlations and SNRs than static conditions.
  • Performance context: Static level 3 achieved MAE 1.96 ± 0.34 beats per minute, heart-rate correlation 0.968 ± 0.079, and waveform correlation 0.626 ± 0.037.A recognizable frequency peak can coexist with waveform power outside the reference fundamental and harmonic bands.
  • Cross-dataset comparison: Beyond Intuition had the highest median skin coverage and SaCo in both evaluations: 0.789 and 0.837 on Static level 3, versus 0.826 and 0.917 on UBFC-rPPG.Secondary attribution rankings differed between the datasets.

6 Conclusion

Skin coverage and SaCo provide information complementary to performance measures rather than serving as proxies for estimation quality. Beyond Intuition was the most useful attribution method across the evaluated scenarios, but its behavior varied by condition and the findings remain descriptive.

  • Conclusion: Within one participant and condition, neither skin coverage nor SaCo was related to clip-level heart-rate error, waveform correlation, or signal-to-noise ratio.186 of 252 coefficients had |ρ| < 0.10, while 28 reached p < 0.05 against 13 expected by chance.
  • Limitations: Clip-pooled analyses answer a different question because pooling measures participants of a condition rather than clips.The pooled analysis produced 146 coefficients at p < 0.05 and reached 0.71, so it should not be read as a within-clip relationship.
  • Conclusion: Beyond Intuition held the highest median refined skin coverage in seven of nine conditions and supplied 14 of the 25 directional relationships.Its attribution rankings were therefore more useful here, although the four methods were neither interchangeable nor independent.
  • Conclusion: Beyond Intuition was the only method whose coverage followed heart-rate error, waveform correlation, and signal-to-noise ratio across the eight scenarios.The three attention-only methods had SaCo values running opposite to each performance measure at that level.
  • Conclusion: In the dimmest condition, Beyond Intuition’s median coverage fell to 0.180 and its median SaCo to −0.178, whereas motion degraded estimates without a comparable attribution drop.This contrast indicates condition-specific attribution behavior rather than a uniform response to degraded estimation.
  • Conclusion: An attribution reveals where the model looks rather than how faithfully its map is ordered or how well the model performs.Attributing to skin alone does not guarantee a good estimate.
  • Conclusion: The Static level 3 results show that high heart-rate correlation can coexist with lower waveform correlation, while mean absolute error hides a clip-level spread of two orders of magnitude.The heart-rate correlation was 0.968 ± 0.079 alongside waveform correlation of 0.626 ± 0.037; the UBFC-rPPG reproduction predicted about twice as accurately, with different secondary attribution rankings.
  • Limitations: The conclusions are descriptive because each condition used one final-epoch checkpoint from a single seed and one architecture, with short 5.12-second windows.The authors call for repeated seeds, additional architectures, common-resolution attribution evaluation, and controlled cross-dataset experiments before causal claims about illumination, generalisation, or clinical validity.

Declarations

The paper reports ethics approval and participant consent, documents data and code-access constraints, and discloses funding, competing interests, author contributions, and generative-AI use.

  • Ethics: The National Cheng Kung University Human Research Ethics Committee approved the experimental procedures, and every participant gave written informed consent.
  • Data availability: UBFC-rPPG is public, whereas NCKU-rPPG is unavailable publicly and participant recordings may be shared by the corresponding author subject to approvals.
  • Code availability: The conversion, training, evaluation, and attribution code is not publicly released but may be provided by the corresponding author on reasonable request.
  • Other declarations: The work lists Taiwan National Science and Technology Council funding, reports no known competing interests, identifies author contributions, and discloses Claude use for editing and analysis-support tasks.

Supplementary Material

The supplementary analyses examine evaluation stability, reference-signal limits, spectral behavior, and relationships between explainability measures and performance. They also compare attribution methods across NCKU-rPPG and UBFC-rPPG, showing that refinement changes magnitudes more than directions.

  • Reference and estimator limits: A 256-frame periodogram resolves 11.7 beats per minute, so a rate within six beats per minute can coexist with a spectrum that is mostly noise.The supplementary results distinguish rate recovery from spectral signal quality.
  • Evaluation-window effects: Lengthening the evaluation window from 5.12 to 61.44 s lowered mean absolute error in every condition, by 1.4 beats per minute on Static level 3 and 6.6 on Bike level 1.The reduction did not erase the separation between static and motion conditions.
  • Reference and estimator limits: The reference photoplethysmogram disagreed with itself by a median of 0.65–1.09 beats per minute when stationary and a mean of 5.49–8.54 when cycling.This bounds how closely a model can agree with the reference on individual clips, especially under cycling.
  • Explainability–performance relationships: 25 of 252 coefficients were directional, with the largest reaching ρ = +0.32; no condition resolved more than five relationships.Beyond Intuition supplied 14 of the 25 directional coefficients, compared with 3, 4, and 4 from the three attention-only methods.
  • Explainability–performance relationships: Within participants, neither skin coverage nor SaCo predicted clip-level model performance in any condition.The measures therefore did not track heart-rate error, waveform correlation, or signal-to-noise ratio at this level.
  • Attribution refinement: Refinement increased raw-attention skin coverage in all eight NCKU-rPPG scenarios, while attention-flow coverage decreased in every scenario.The corresponding UBFC-rPPG changes had the same directions but larger magnitudes: +0.224, −0.117, and −0.167.

S7. Heart-rate estimation settings in the remote photoplethysmography literature since 2020

The literature review finds substantial variation and incomplete reporting in heart-rate estimation settings for remote photoplethysmography. Window length, spectral settings, overlap, and estimator choices materially affect comparability, motivating explicit reporting and controlled variation.

  • Evaluation windows: Window lengths range from 5 s to 60 s, while 30 s and 10 s are the two commonest choices.In these data, changing the window from 10.24 to 30.72 s changed mean absolute error by 0.23 to 1.63 beats per minute across eight conditions.
  • Spectral settings: Only nine of 58 articles state the transform length or window function, despite a 160-frame clip resolving no better than 11 beats per minute without zero-padding.Reported spectral settings vary widely, including different transform lengths and zero-padding choices.
  • Frame rate: Fifty of 58 articles process video at 30 Hz, whereas the present recordings use 50 Hz, making equal frame-count windows shorter in seconds here.The review therefore reports both frame counts and durations.
  • Reporting completeness: Seventeen of 58 articles state no heart-rate window, 23 state no overlap, and four do not identify the spectral method.These omissions make many published estimates difficult to reproduce or compare.
  • Comparability: A published mean absolute error is not comparable across studies without the window, overlap, spectral settings, and estimator, which cannot be recovered for close to a third of the literature.The review records these settings explicitly and varies the most influential one in Table 19.
Loading 2609.03663v1…