Automatic denoising and differentiation based on Savitzky-Golay filtering and Homogeneous Differentiators for attractor reconstruction via differential embedding
Uros Sutulovic, Daniele Proverbio, Rami Katz, Giulia Giordano
q-bio.QM · eess.SP
Quick verdict
SHADED convincingly removes manual HD/SG parameter tuning for differential embedding, but validation is largely qualitative and lacks quantitative baselines.
Summary
The paper introduces SHADED, a fully automatic pipeline that chains Homogeneous Differentiator (HD) and Savitzky-Golay (SG) blocks to denoise a scalar time series and estimate derivatives up to arbitrary order, enabling attractor reconstruction via differential embedding. The HD gain is selected by simulated-annealing minimisation of a residual standard-deviation cost, and SG windows by a cost balancing chattering suppression against low-frequency signal preservation, with an optional Welch-PSD frequency-persistence adjustment. Across Lorenz and three neuroscience models under additive and multiplicative white Gaussian noise, LTspice transistor circuits with internally injected noise, and PPG/ABP recordings from three patients (UQ Vital Signs Dataset), the automatic parameters land close to inaccessible RMSE-optimal values and reconstructed embeddings are visually faithful; ABP attractors show a consistent double-loop structure. Main limitations: accuracy assessment is mostly visual/geometric rather than quantitative; the frequency-separation assumption is heuristic; several hyperparameters (SG degree d=2, cost weight lambda=0.5, gain perturbation epsilon=0.05, 5% transient trimming) remain fixed defaults; and there is no benchmarking against alternative automatic differentiation or denoising pipelines.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SHADED automatically infers all HD gains and SG window sizes from noisy data alone, eliminating manual tuning that previously limited HD-SG differential embedding [27].
- On ground-truth systems, automatically selected parameters achieve derivative errors and reconstructions comparable to RMSE-optimal parameters that require access to the noise-free signal.
- The resulting differential embeddings faithfully preserve attractor geometry across neuronal models, transistor-based chaotic circuits, and real cardiovascular recordings under varying SNRs and noise types.
- The optional frequency-persistence adjustment prevents oversmoothing where the frequency-separation assumption fails, preserving fine attractor features (u(2) range expanded by factors of roughly 2-7).
- Differential embeddings of ABP/PPG signals reveal a reproducible double-loop structure plausibly associated with distinct cardiac-cycle phases.
Strengths
Solves a real, well-motivated open problem
The paper directly addresses the manual-tuning bottleneck of the authors' prior HD-SG method [27], and frames automatic SG-window and HD-gain selection as a long-standing open problem [39-42]; the residual-based gain cost (Eq. 6) and the chattering/signal-preservation SG cost (Eq. 10) are concrete, implementable data-driven criteria.
Automatic parameters match inaccessible optimal values on ground-truth systems
In the Lorenz example (Fig. 2), the inferred gain L̂=540 sits close to the RMSE-optimal L*=553, and the estimated SG window T̂ approaches T* (Fig. 3); the staircase recovers noise-free attractor geometry for Lorenz, Hindmarsh-Rose, FitzHugh-Nagumo and Jansen-Rit under both additive and multiplicative noise (Fig. 4).
Thoughtful engineering and honest handling of no-ground-truth regimes
Breadth of validation domains
Weaknesses
Evaluation is predominantly qualitative
Heuristic cost with a strong frequency-separation assumption
'Fully automatic' claim is diluted by many fixed hyperparameters
Clinical interpretation rests on three patients and unvalidated associations
No comparison against competing automatic methods
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SHADED and what problem does it solve?
SHADED (Savitzky-Golay and Homogeneous-differentiator based Automatic DEnoising and Differentiation) is a signal-processing pipeline that takes a noisy scalar time series and automatically produces a denoised signal plus its successive derivatives. Its purpose is attractor reconstruction via differential embedding, which previously required laborious manual tuning of differentiator gains and filter window sizes.
How does SHADED choose its parameters automatically?
For each stage, the differentiator gain is selected by minimising the standard deviation of the residual (input minus output) using simulated annealing, and the Savitzky-Golay window is chosen by minimising a cost that balances removing high-frequency chattering against preserving the low-frequency signal content; an optional spectral check keeps the smallest window that still removes the dominant chattering frequency.
What data was SHADED tested on?
It was tested on simulated data from the Lorenz system and three neuroscience models (Hindmarsh-Rose, FitzHugh-Nagumo, Jansen-Rit) with additive and multiplicative Gaussian noise, on LTspice simulations of chaotic transistor circuits with internally injected noise, and on real photoplethysmography (PPG) and arterial blood pressure (ABP) recordings from three patients in the University of Queensland Vital Signs Dataset.
Does SHADED only remove noise, or does it also compute derivatives?
It does both: each cascade stage yields a denoised signal, and successive stages estimate higher-order derivatives up to a desired order. The denoised derivatives are useful in their own right (e.g., for model discovery or landmark detection) and serve as coordinates for reconstructing an attractor via differential embedding.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The embedded text contains only the main text; the Supplementary Material (algorithms' details, sensitivity studies S1-S34) is referenced but not provided, so several claims could not be checked directly.
- Figures are represented only through captions and axis-label fragments in the extraction, so visual claims about reconstruction quality could not be independently assessed.
- Reconstruction-accuracy evidence in the main text is mostly qualitative, which limits how strongly this review can endorse the accuracy claims regardless of extraction quality.
- Reported quantitative values (gains, windows, runtimes) were transcribed from figure captions and text; figures themselves were not analyzable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.