PaperlayerView exact revision on arXiv

Source-linked AI summary

Cyclostationary Phase Conditioning for Medical Time Series Diffusion

Samuel Ruiperez-Campillo, Michele Copetti, Jorge da Silva Goncalves, Sonia Laguna, Thomas Hofmann, Julia E. Vogt

arXiv:2609.34965v2cs.LGcs.AIeess.SP

TL;DR

Physiological time series have cycle-dependent statistics, yet existing diffusion restorers condition only on corrupted observations. PTSDIFF adds a shift-covariant wavelet representation and learned phase conditioning, alongside a training-free phase-utility index and antithetic sampling. Across modalities, the paper reports improved restoration and downstream performance, with antithetic sampling matching or exceeding ten trajectories using fivefold fewer network evaluations.

  • Problem

    Physiological recordings are corrupted, and existing diffusion restorers condition only on observations rather than explicitly modeling their approximately periodic, cycle-dependent statistics.

    Source: §1 · p. 1

  • Method

    PTSDIFF combines shift-covariant wavelet diffusion and learned dense phase conditioning with a training-free cyclostationarity index and antithetic reverse sampling.

    Source: §1 · p. 2

  • Results

    Across four modalities, PTSDIFF achieves state-of-the-art restoration and improves downstream-task performance; one antithetic pair matches or exceeds ten independent trajectories with fivefold fewer network evaluations.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    Phase conditioning’s restoration benefit increases with cyclostationary structure, and the index estimates when phase conditioning is likely to help before training.

    Source: §1 · p. 2, §6 · p. 9

    A clean-looking output may still be diagnostically wrong; distortion metrics alone do not establish diagnostic fidelity.

    Source: ETHICS STATEMENT · p. 11

Abstract

from arXiv · show

Many physiological time series, such as cardiac and brain recordings, exhibit cyclostationarity: their statistics vary periodically with an underlying cycle phase. Corruption from motion, poor contact, and physiological interference obscures morphology needed for diagnosis, making signal restoration essential. Existing diffusion approaches condition on corrupted observations alone and must learn cyclic structure implicitly. We instead propose two inductive biases which encode cyclostationarity: a shift-covariant wavelet representation and dense per-sample phase conditioning inferred from the corrupted input. We further introduce a training-free cyclostationarity index that quantifies phase structure and predicts when phase conditioning will help. Finally, we propose antithetic coupling of reverse trajectories to reduce sampling variance while achieving comparable performance with fivefold fewer network evaluations. Across modalities, our results show that explicitly encoding measurable cyclic structure improves physiological time-series restoration.

1 INTRODUCTION

Physiological recordings contain recurring cycle structure, but existing diffusion restorers condition only on corrupted observations. PTSDIFF makes cycle structure explicit and introduces measures and sampling strategies intended to improve restoration efficiently.

1 INTRODUCTION

  • Physiological recordings support diagnosis and monitoring but are corrupted by movement, poor contact, and interference from other physiological sources.Their statistics vary approximately periodically with latent cycle phase, a structure existing diffusion restorers leave implicit when conditioning only on corrupted observations.Source: §1 · p. 1
  • PTSDIFF combines a shift-covariant wavelet representation with dense phase conditioning inferred from the corrupted input to encode cycle structure explicitly.The representation addresses arbitrary window-start position, while phase conditioning provides position within the cycle.Source: §1 · p. 2
  • The cyclostationarity index estimates from signal structure when phase conditioning is likely to help, avoiding unnecessary model complexity and training when cycle structure is weak.The paper also finds phase-estimation quality matters: a classical detector performs no better than omitting phase conditioning, while a learned estimator improves restoration.Source: §1 · p. 2
  • PTSDIFF achieves state-of-the-art restoration across four modalities and improves downstream-task performance; one antithetic pair matches or exceeds ten independent trajectories with fivefold fewer network evaluations.The framework combines shift-covariant wavelet diffusion and cycle-phase conditioning with antithetic reverse sampling.Source: §1 · p. 2

2 RELATED WORK

Prior restoration work spans classical pipelines, learned deterministic models, and diffusion approaches, while time-series diffusion explores alternative representations. Existing periodicity work uses cycles but does not estimate how much a signal may benefit from them.

2 RELATED WORK

  • Classical restoration remains standard, while learned deterministic restorers outperform it but learn a single map and offer no control over estimator variance.Classical methods can degrade under non-stationary artifacts and suppress low-amplitude diagnostic morphology.Source: §2 · pp. 2–3
  • Generative restorers model time-series modalities, but general-purpose samplers target inverse problems and accelerated solvers reduce steps rather than estimator variance at fixed NFE.The related-work discussion identifies fixed-NFE variance reduction as a distinct target.Source: §2 · p. 3
  • Time-series diffusion has used spectral and time-frequency representations; PTSDIFF adopts the stationary wavelet transform for shift-covariance rather than spectral properties.MECG-E uses the STFT, while TFCDiff uses the DCT.Source: §2 · p. 3
  • Prior physiological models exploit repetition, but do not determine how much a given signal should benefit; PTSDIFF also evaluates antithetic coupling at matched NFE.Antithetic variates were introduced into conditional diffusion sampling for physiological time series by Ruip´erez-Campillo et al. (2026a).Source: §2 · p. 3

3 METHOD: PTSDIFF

PTSDIFF incorporates cyclostationary structure into diffusion through a shift-covariant frame and explicit phase conditioning, then estimates phase utility and reduces inference-time sampling variance.

3.1 PRELIMINARIES

  • The method assumes wide-sense cyclostationarity: signal statistics depend on latent cycle phase rather than absolute time, allowing successive cycles to vary in length and morphology.This assumption is weaker than periodicity and stronger than stationarity.Source: §3.1 · p. 3

3.2 DIFFUSION IN A SHIFT-COVARIANT FRAME

  • Diffusion uses the stationary wavelet transform, whose undecimated channels preserve temporal resolution and remain aligned under shifts for length-preserving denoising.With four levels, the representation has four detail channels and one approximation channel; phase channels can be transformed and concatenated without resampling.Source: §3.2 · p. 4

3.3 PHASE CONDITIONING

  • A learned shift-equivariant encoder estimates a dense phase field from corrupted input, then its circle-embedded phase channels are frame-transformed and concatenated with diffusion inputs.For ECG, a reliable event detector warm-starts the encoder; other modalities are trained without warm-starting.Source: §3.3 · p. 5

3.4 MEASURING CYCLE STRUCTURE

  • The cyclostationarity index Π is the fraction of signal variance explained by cycle phase and can be estimated from clean data before model training.Phase-aligned averaging estimates the phase-dependent mean; Π exactly quantifies the phase-conditioning MMSE reduction without observations, but is not a certificate when y is observed.Source: §3.4 · p. 5

3.5 LEARNING OBJECTIVE

  • The learning objective addresses the mismatch between frame-domain noise-prediction error and time-domain reconstruction error with reconstruction and temporal-slope penalties.The slope penalty encourages sharp transients and is restricted to low noise; the reconstruction and morphology weights are 0.3 and 0.1, respectively.Source: §3.5 · p. 6

3.6 ANTITHETIC REVERSE ESTIMATION

  • At matched network-evaluation budgets, antithetic sampling averages pairs of trajectories driven by opposite noise while remaining unbiased for the model’s conditional mean.When pair correlation is negative, coupling reduces the compute-controllable variance term; it is no worse than independent sampling to first order when correlation is near zero.Source: §3.6 · p. 6

4 EXPERIMENTAL SETUP

The experiments assess restoration across four physiological modalities using modality-specific datasets and baselines. Performance is measured with restoration metrics and bootstrap intervals under a specified training and sampling protocol.

4 EXPERIMENTAL SETUP

  • Table 1 compares restoration across four modalities against modality-specific baselines, reporting mean ± standard deviation over three training seeds.Best and second-best values are marked within each modality; modalities are ordered by phase-shuffle margin.Source: Table 1 · p. 7
  • The study evaluates PPG, ECG, EMG, and EEG, including held-out datasets for testing distribution shift in each modality.ECG also uses held-out MIT-BIH Arrhythmia and QTDB datasets; the PPG, EMG, and EEG evaluations each include one held-out dataset.Source: §4 · p. 7
  • Each modality is compared with its strongest published restoration methods, retrained on identical splits and preprocessing.The modality-specific baseline sets include DDAE for PPG; TCDAE, DeScoD-ECG, and TFCDiff for ECG; SDEMG and FCN for EMG; and EEGDfus and EEGIFNet for EEG.Source: §4 · p. 7
  • Restoration is evaluated using ∆SNR, PRD, and Pearson correlation, averaged across signals with 95% bootstrap confidence intervals.Paired baseline comparisons use one-sided Wilcoxon signed-rank tests with Holm–Bonferroni correction.Source: §4 · p. 7

5 RESULTS

PTSDIFF improves physiological-signal restoration and downstream utility across modalities, while phase-conditioning benefits vary with cycle structure and phase-estimation quality. Antithetic sampling improves the quality–compute trade-off.

Restoration performance

  • PTSDIFF-AV-10 achieves the best ∆SNR on all four modalities and the best PRD and CC on three; under dataset shift, it leads on PPG, ECG, and EMG, but not EEG.EEGIFNet retains the best PRD and CC on EEG.Source: §5 · p. 8
  • PTSDIFF matches or exceeds the strongest restoration baseline on every downstream task across four modalities.On ECG, it recovers most of the macro-AUROC and all abnormal-beat sensitivity lost to noise, with the best F1 among restoration methods.Source: Table 2 · p. 8, §5 · p. 8

Phase conditioning and cyclostationarity

  • r = 0.981 across four modalities links phase-shuffle sensitivity to PTSDIFF’s margin over the strongest baseline; synthetic ∆SNR gain rises with cyclostationarity index Π (r = 0.85).The synthetic study varies cyclic variance share and cycle-length jitter while holding other factors fixed.Source: §5 · p. 8, Figure 3 · p. 8

Further Analyses

  • Each component improves ∆SNR in three of four modalities, and phase and context conditioning are complementary on ECG.Time-domain auxiliary losses help most on EEG; the shift-covariant frame contributes comparably on ECG, EMG, and EEG, but slightly hurts PPG.Source: §5 · pp. 8–9

Phase conditioning and cyclostationarity

  • On ECG, the learned phase encoder improves every metric over the phase-free model, while phase from a classical detector performs on par with omitting phase.The ECG ablation uses PTB-XL at SNR −2.0 dB with one antithetic pair and a single seed.Source: §5 · p. 8, Table 3 · p. 9

Sampling efficiency

  • At equal NFE, PTSDIFF-AV-10 improves ∆SNR and PRD over PTSDIFF-MC-10 in every modality.On ECG, one antithetic pair outperforms ten independent trajectories with five times fewer NFEs; further pairs add little as model bias dominates.Source: §5 · p. 9

6 CONCLUSION

The conclusion presents PTSDIFF as a cyclostationary-aware restoration framework and reports benefits across four real-world modalities, while identifying broader data and multichannel evaluation as future work.

6 CONCLUSION

  • A single antithetic pair outperforms ten independent ECG trajectories with 5× fewer network evaluations, and wall-clock time decreases similarly.Source: Figure 4 · p. 9, §6 · p. 9
  • PTSDIFF outperforms prior restoration baselines across four real-world modalities and datasets.Source: §6 · p. 9
  • PTSDIFF combines a shift-covariant frame representation, learned dense phase field, training-free cyclostationarity index, and antithetic reverse sampling.Source: §6 · p. 9
  • Phase-conditioning gains increase with cyclostationary structure, while learned phase estimation improves restoration where analytical phase from corrupted input does not.The index separates cardiac modalities from less cyclic ones, and its autocorrelation proxy orders modalities similarly to ablation gains.Source: §6 · p. 9
  • The relationship between cyclostationarity and restoration benefit needs testing on broader datasets and naturally occurring artifacts; current analyses focus on single-channel settings.The approach also remains more expensive than single-pass restoration despite antithetic sampling.Source: §6 · pp. 9–10
  • Distortion metrics do not establish diagnostic fidelity, and synthetic corruption from few artifact sources leaves robustness to real ambulatory settings unverified.The authors call for downstream task validation on the population of interest before diagnostic deployment.Source: ETHICS STATEMENT · p. 11

A EXTENDED RELATED WORK

  • Appendix A expands the main text’s idea-organized prior work with modality-specific ECG, PPG, EEG, and EMG restoration histories and broader reconstruction contexts.Source: A EXTENDED RELATED WORK · p. 19

B.1 PROOF OF PROPOSITION 1 (PHASE-EXPLAINED VARIANCE)

  • Π exactly quantifies unconditional MMSE reduction from phase, but need not bound phase’s added benefit once observation y is available.Phase-dependent conditional variance can make the observed-data benefit positive even when Π = 0.Source: §B.1 · p. 20

B.2 PROOF OF PROPOSITION 2 (MATCHED-NFE VARIANCE)

Antithetic pairing preserves the conditional-mean estimate while reducing sampling variance when pair correlation is negative, with gains limited by model bias. The cyclostationarity index is estimated from phase-aligned clean cycles, while an autocorrelation proxy can screen signals when event detection is unavailable.

B.2 PROOF OF PROPOSITION 2 (MATCHED-NFE VARIANCE)

  • At matched network evaluations, antithetic averaging has variance ratio 1 + ρ(τ) relative to independent averaging, so it improves variance exactly when ρ(τ) < 0.The pair estimator is unbiased for the same conditional mean estimated by independent averaging.Source: §B.2 · p. 20
  • At matched network evaluations, antithetic coupling reduces error in the conditional mean, but gains decay as 1/K and saturate at model bias.The reduction concerns distortion, not the learned distribution or individual-sample quality.Source: §B.2 · p. 20
  • PTSDIFF’s variance-weighted pair correlation is negative in every modality, never exceeding −0.89 along the reverse trajectory and relaxing near t = 0.This agrees closely with reported conditional-diffusion pair correlations of −0.98 to −0.88 through most of the trajectory.Source: §B.2 · p. 21

B.3 ESTIMATING THE CYCLOSTATIONARITY INDEX

  • The cyclostationarity index is estimated from clean signals by detecting cycle events, resampling each cycle to 100 phase bins, then averaging and normalizing.Cycles outside the physiological range are discarded; amplitude-normalizing each cycle makes the index measure shape consistency rather than amplitude stability.Source: §B.3 · p. 21
  • At small sample counts, the index estimate is upward-biased by residual noise in the mean cycle; the correction changes it by less than 0.01 when N ≥ 200.The correction subtracts within-bin variance divided by N.Source: §B.3 · p. 21
  • When reliable event detection is unavailable, peak normalized autocorrelation within plausible cycle lags provides a one-pass screening proxy, but measures single-lag self-similarity rather than the index’s variance decomposition.It is not the quantity bounded by Proposition 1.Source: §B.3 · p. 21

C.1 PER-MODALITY PREPROCESSING

The appendix documents modality-specific preprocessing, shared model and baseline implementations, metrics, and statistical procedures. On ECG, PTSDIFF is best across the reported metrics and datasets except for a tie on QTDB MAD.

C.1 PER-MODALITY PREPROCESSING

  • For ECG, EMG, and EEG, corruption noise comes from held-out records or record halves disjoint from training noise, whereas PPG training and test noise use the same segments.Dataset-specific splits and corruption protocols follow established benchmarks.Source: §C.1 · p. 22
  • The modality protocols define distinct data and corruption setups, including EEG–EOG mixtures, ECG noise from NSTDB, PPG motion-artifact mixtures, and EMG contaminated with ECG interference.The appendix specifies held-out datasets and additional preprocessing for each modality.Source: §C.1 · pp. 22–23

C.2 ARCHITECTURE AND TRAINING DETAILS

  • The denoiser is a three-level 1D U-Net conditioned on concatenated frame-transformed inputs and dense phase fields, with resolution-specific FiLM embeddings.For ECG, the phase encoder is warm-started using an analytical clean-signal detector but does not use that detector at inference; EEG also has a configuration exception.Source: §C.2 · pp. 23–24, Table 5 · p. 24

C.3 BASELINE IMPLEMENTATION DETAILS

  • Learned baselines are retrained on shared prepared tensors and splits; modality-specific implementations retain published architectures and objectives, with documented adaptations for data interfaces or compute.Examples include the PPG DDAE reimplementation and EMG SDEMG batch-size and gradient-accumulation changes.Source: §C.3 · pp. 24–25

C.4 EVALUATION METRICS

  • Evaluation reports three metrics across modalities plus literature-specific benchmark metrics, computed per test sample and averaged.For EEG, the spectral metric uses the power spectral density; EMG metrics include mean rectified value and mean frequency.Source: §C.4 · p. 25, Table 6 · p. 26

C.5 STATISTICAL EVALUATION PROCEDURE

  • Paired PTSDIFF–baseline differences are tested with one-sided Wilcoxon signed-rank tests and Holm–Bonferroni correction; every reported comparison favors PTSDIFF and remains significant.The procedure uses paired samples by test-set index and α = 0.05.Source: §C.5 · p. 26

D.1 FULL ECG RESULTS ACROSS DATASETS

  • PTSDIFF is best on every reported ECG metric and dataset except MAD on QTDB, where it ties TFCDiff.Table 7 reports the full comparison across PTB-XL, QTDB, and MIT-BIH.Source: §D.1 · p. 26

D.2 RESULTS STRATIFIED BY CORRUPTION LEVEL

PTSDIFF’s advantage generally persists across corruption levels, with the ECG results strongest across the reported bins and benchmark-specific metrics. EMG amplitude and mean-frequency errors are more mixed under heavy corruption.

D.2 RESULTS STRATIFIED BY CORRUPTION LEVEL

  • Figure 6 groups restoration results by each modality’s corruption grid, with y-intercepts marking average performance per model and metric and arrows indicating the better direction.Figure 7 repeats the stratification using each benchmark’s own metrics.Source: Figure 6 · p. 27, Figure 7 · p. 28
  • Across corruption levels, benchmark-metric ordering follows the main metrics on PPG and EEG, while EMG ARV and mean-frequency error are more mixed under heavy corruption.Figure 7 presents the modality-specific benchmark metrics; the passage associates the EMG pattern with mild smoothing from trajectory averaging.Source: §D.2 · p. 26, Figure 7 · p. 28
  • PTSDIFF is best on every ECG metric in every corruption bin on every dataset, except MAD in the 0.2–0.6 and 0.6–1.0 bins.Tables 8–10 stratify the ECG comparisons by corruption scaling factor λ.Source: §D.2 · p. 26, Table 8 · p. 29, Table 9 · p. 30

D.3 CROSS-DATASET GENERALIZATION

Without retraining, PTSDIFF-AV-10 leads ECG comparisons on MIT-BIH and QTDB and improves over the strongest held-out baselines on PPG and EMG. EEG cross-dataset results are more limited, while paired ECG comparisons remain significant after correction.

D.3 CROSS-DATASET GENERALIZATION

  • On EEGdenoiseNet-to-EEGMMIDB transfer, EEGIFNet remains ahead on all three metrics by 0.37 dB ∆SNR, while PTSDIFF exceeds EEGDfus by 2.02 dB ∆SNR, 13.54 PRD points, and 0.105 CC.All restorers perform markedly worse than in-domain, with clean EEG below 4 Hz removed substantially.Source: §D.3 · p. 32
  • On ECG, the diffusion baselines lose their advantage over deterministic TCDAE on MIT-BIH, while PTSDIFF does not; the authors interpret this as evidence for beat-local rather than global-rhythm conditioning.The authors relate the distinction to arrhythmic beats violating a global period while retaining a well-defined local phase field.Source: §D.3 · p. 32
  • Without dataset-specific retraining, PTSDIFF-AV-10 achieves the highest ∆SNR and lowest PRD on MIT-BIH and QTDB; it also improves on the strongest PPG and EMG baselines by 2.08 and 1.49 dB ∆SNR, respectively.Table 11 summarizes train-on-one-dataset, test-on-held-out-dataset comparisons against the strongest baseline on each held-out set.Source: Table 11 · p. 32, Figure 8 · p. 32, §D.3 · p. 32

D.4 PAIRED STATISTICAL COMPARISON

  • PTSDIFF-AV-10 is significantly better than every learned and classical baseline on all three main metrics and all three ECG datasets after Holm–Bonferroni correction.Table 12 reports the paired tests behind these ECG significance claims.Source: §D.4 · p. 32

D.5 EXTENDED DOWNSTREAM EVALUATION

The section evaluates whether restoration quality carries through to downstream tasks, using classification on ECG and task-specific utility measures across modalities. The reported classification gains are not uniform across operating points.

D.5 EXTENDED DOWNSTREAM EVALUATION

  • AUROC gains are not purely a ranking effect, but PTSDIFF is not uniformly best on sensitivity and specificity, so downstream benefit depends on the operating point.The single-lead classifiers score below 12-lead references, but that does not affect comparisons across restoration methods.Source: §D.5 · p. 32
  • PTB-XL and MIT-BIH are evaluated with per-class AUROC and binary beat classification, while PTB-XL also reports sensitivity, specificity, and F1.The tables report point estimates with 95% bootstrap confidence intervals.Source: §D.5 · p. 32, Table 13 · p. 34, Table 14 · p. 34
  • The cross-modal utility evaluation uses heart-rate prediction for PPG, muscle-activation detection for EMG, and EEG band-power error alongside ECG beat classification.The evaluation reports MAE, Pearson r and R2 for PPG; balanced accuracy for EMG; and relative error for EEG band power.Source: §D.5 · p. 34, Table 16 · p. 35

D.6 COMPONENT ABLATION ACROSS MODALITIES

Component ablations show modality-dependent contributions from phase conditioning and the wavelet frame. Across datasets and synthetic signals, stronger measured cyclostationarity aligns with greater benefit from phase conditioning, while frame choice and decomposition depth matter for ECG performance.

D.6 COMPONENT ABLATION ACROSS MODALITIES

  • Removing phase costs most on PPG (−1.82 dB) and ECG (−0.64 dB), while removing the frame costs 0.5–0.7 dB on ECG, EMG, and EEG but slightly helps PPG.The ablation is reported relative to the full model; the analytical-phase row is included only for ECG.Source: §D.6 · p. 35, Table 17 · p. 35

D.7 CYCLOSTATIONARITY INDEX ACROSS DATASETS

  • Mean autocorrelation proxy ΠAC orders modalities PPG (0.74) > ECG (0.71) > EEG (0.37) > EMG (0.10), matching the order of phase-ablation costs.The detector-based ˆΠ is reported for ECG only.Source: §D.7 · p. 35, Table 18 · p. 36

D.8 SYNTHETIC CYCLOSTATIONARITY SWEEP

  • Measured cyclostationarity predicts larger gains from periodicity conditioning in the synthetic sweep (r = 0.85 overall; r = 0.82–0.88 within fixed-share groups).The sweep varies cyclic variance share and cycle-length jitter; its curves are descriptive and do not imply a particular functional relationship.Source: §D.8 · p. 36

D.9 FRAME AND OBJECTIVE SELECTION

  • Across the frame sweep, ∆SNR ranges from 14.58 (coif3, J = 8) to 15.98 (sym4, J = 4); performance degrades for J > 4 across wavelet families.The authors attribute the deeper-level decline to the approximation band falling below the signal’s fundamental, so the coarsest channel no longer carries cycle-scale morphology.Source: §D.9 · p. 36

D.10 PHASE-SHUFFLE EXPERIMENT

The appendix tests phase reliance, antithetic sampling, per-sample performance distributions, and qualitative reconstructions. Antithetic sampling reduces compute while retaining restoration quality, and the qualitative examples are illustrative rather than evidence.

D.10 PHASE-SHUFFLE EXPERIMENT

  • Shuffling phase fields across samples lowers ∆SNR in all four modalities, confirming use of the conditioning channel; the shuffled-field penalty exceeds the ECG phase-ablation gap.The reported ECG drops are +4.14 dB for shuffled phase versus 0.64 dB for ablated phase.Source: §D.10 · pp. 36, 38, Figure 9 · p. 38

D.11 ANTITHETIC SAMPLING

  • One antithetic pair already exceeds ten independent trajectories on ECG, with five times fewer network evaluations; measured pair correlation stays negative throughout the reverse process.Further pairs add little once model bias dominates, and sampler swaps also help DeScoD-ECG and TFCDiff without retraining.Source: §D.11 · p. 38, Table 21 · p. 38, Figure 11 · p. 39

D.12 PER-SAMPLE DISTRIBUTION

  • PTB-XL per-sample ∆SNR improvements reflect a shifted distribution rather than a heavy tail, with distributions also stratified by corruption scaling factor λ.The figure reports the overall distribution and strata; box plots show IQR, median, 95% CI notches, whiskers, and outliers.Source: §D.12 · p. 40, Figure 12 · p. 40

D.13 THE FRAME DECOMPOSITION

  • The four-level sym4 frame keeps one approximation and four detail channels at the input length; baseline wander concentrates in a4, while motion and muscle artifacts spread across detail channels.The prose identifies d1 and d2 as carrying predominantly artifact energy.Source: §D.13 · p. 40, Figure 13 · p. 41

D.14 QUALITATIVE RESULTS

  • Figures 14–16 show single-sample reconstructions for heavy and light ECG corruption and for PPG, EMG, and EEG; these examples are illustrative, not evidence.The non-ECG examples cover light and heavy corruption levels for each modality.Source: §D.14 · p. 40, Figure 14 · p. 41, Figure 15 · p. 42

Related papers

Loading 2609.34965v2…