Learned denoisers for heart recordings work by squeezing a noisy signal through a narrow internal representation. Random sensor noise cannot pass through that narrow channel, so it is discarded. Rare heartbeat types may be discarded for the same reason. Fusion beats, unusual ventricular beats and other minority morphologies are, by definition, far from the shapes the denoiser saw most often during training. They also tend to live in the same sharp, high-frequency part of the signal as muscle noise. This proposal tests a simple and uncomfortable hypothesis: a denoising autoencoder trained on a realistically imbalanced ECG archive removes noise partly by "normalizing" rare beats toward common ones, and this damage is concentrated at specific temporal scales.
To measure the effect and to correct it, we build a denoiser that grows from coarse to fine temporal resolution. Its output is an explicit sum of per-scale contributions. That structure lets us show exactly what was removed at each scale for each beat. It also lets us apply a minority-aware fidelity objective only at the scales where rare morphology and noise overlap. The deliverable has three parts. The first is a per-class, per-scale account of what current denoisers destroy. The second is a denoiser that keeps minority-class recall without giving up noise suppression. The third is a "what did the denoiser throw away" display that a clinician or engineer can inspect for any beat.
Core Claim
The claim has two halves, and either half can fail independently. The empirical half says the following. At matched input noise levels, a fixed-resolution denoising autoencoder distorts minority AAMI classes (the standard five heartbeat groups N, S, V, F, Q) more than normal beats. It does so measurably, both in waveform error and in downstream minority recall, and the excess distortion sits mainly in the finest temporal bands. The methodological half says that a progressively grown, additively decomposed denoiser can fix this by applying class-balanced reconstruction weighting only at those fine bands. The fix should recover minority recall better than balancing the whole fixed-resolution denoiser. The scale decomposition also gives a faithful-by-construction view of the removed signal, which post-hoc saliency on a monolithic denoiser cannot give.
Relation to Prior Work
Three papers motivate the premise and supply the starting point. The fully convolutional denoising autoencoder of Noise Reduction in ECG Signals Using Fully Convolutional Denoising Autoencoders compresses a 1024-sample window into a 32-sample bottleneck. On MIT-BIH beats mixed with Noise Stress Test Database noise, it reports a signal-to-noise improvement of 15.49 dB at −1 dB input SNR. That paper reports aggregate fidelity only, never per arrhythmia class. Measuring fidelity per class is the gap we fill.
Low-dimensional Denoising Embedding Transformer for ECG Classification replaces learned embeddings with an 18-dimensional wavelet-and-Fourier feature set and rebalances classes with SMOTE (synthetic minority oversampling by interpolating between neighboring examples). It reports 98.39% average recall. SMOTE operates after the denoising embedding is fixed, so it cannot recover morphology the embedding has already discarded.
The research file's own synthesis, drawing on Mitigating Class Imbalance in Clinical Machine Learning, names this denoising-versus-minority tension as open. No paper in the file measures it.
Two lines of work supply the method. Progressive growing from Progressive Growing of GANs for Improved Quality, Stability, and Variation trains at low resolution first and then fades in higher-resolution layers through a blending weight. Improving trust and confidence in medical skin lesion diagnosis through explainable deep learning showed that the same trick (14×14 to 224×224) stabilized an autoencoder over rare skin-lesion classes. Both work on 2D images. We use progressive growth on 1D signals for a different reason. We are not after generative stability. We want an additive per-scale decomposition in which each band's contribution is learned and inspectable separately.
Two lines of work constrain the interpretability design. Clinically meaningful interpretability of an AI model for ECG classification showed that Jacobian saliency on ECGs matches electrophysiologist annotations (correlation 0.94 for premature ventricular contractions). This makes it the right saliency competitor, but it explains a classifier, not what a denoiser removed. Climbing the Ladder of Interpretability with Counterfactual Concept Bottleneck Models argues that counterfactuals ("what minimal change flips the decision") beat heatmaps, with 99% versus 66% validity on CUB. We borrow that idea in a weak form: we swap one scale's contribution and check whether the diagnosis flips.
TolerantECG handles noise through distillation from a noise-expert teacher. It sidesteps explicit reconstruction, and it reports 0.979 AP on 2-lead MIT-BIH. It is a strong robust classifier, but it does not show users what was removed. Contrastive pretraining such as Self-supervised representation learning from 12-lead ECG data likewise gains noise invariance without a reconstructed waveform.
Research Program
The denoiser represents its output as a sum over dyadic temporal scales:
$$\hat{x} = \sum_{s=0}^{S} U_s\big(r_s\big),$$
The symbols are defined as follows:
- $\hat{x}$ is the denoised beat window at full sampling rate.
- $s=0$ is the coarsest scale, roughly 360 Hz downsampled by $2^S$, where rhythm context and the slow ST–T shape live.
- $r_s$ is the residual waveform predicted by stage $s$ at that stage's sampling rate.
- $U_s$ is fixed linear upsampling to full rate.
Training starts with stage 0 alone. Each finer stage is then faded in with a blending weight that ramps from 0 to 1, as in progressive GANs, while earlier stages continue training at a reduced learning rate. Every stage shares the encoder style of the fully convolutional autoencoder, with strided convolutions and no pooling, so the comparison with that baseline is architectural rather than a matter of capacity.
The minority-aware objective weights reconstruction error per scale and per class:
$$\mathcal{L} = \sum_{s} \sum_{c} w_{s,c}\, \frac{1}{|B_c|}\sum_{i \in B_c} \big\| P_s(x_i) - P_s(\hat{x}_i) \big\|^2 .$$
Here $x_i$ is the clean reference beat, $B_c$ is the set of beats of class $c$ in the batch, and $P_s$ is a band projection onto scale $s$. The weights $w_{s,c}$ equal 1 at coarse scales and inverse class frequency at the finest $k$ scales. The integer $k$ is chosen from the audit, not tuned on test data.
Class labels are needed only during training. At inference the denoiser is label-free.
The program has three stages. The first is the audit. We train the fixed-resolution autoencoder and the unweighted progressive model on MIT-BIH beats mixed with Noise Stress Test Database baseline wander, muscle artifact and electrode motion at input SNRs from −6 to 24 dB. We then report waveform error per AAMI class and per scale band, energy removed inside the QRS window per class, and downstream per-class recall from a fixed classifier trained on clean beats.
The second stage is the fix: the weighted progressive model above.
The third stage is the removed-signal display. For any beat we show $x_{\text{noisy}} - \hat{x}$ split into its scale components. We also flag beats where the removed fine-scale component correlates with the class-mean QRS template of a minority class, because such a correlation suggests the denoiser may have erased evidence.
As a faithfulness check, we run scale-swap counterfactuals. We replace stage $s$'s contribution for a beat with the class-mean contribution of another class and ask whether the downstream classifier flips. This identifies which scale carries each class's diagnostic evidence, with no gradient approximation involved.
All MIT-BIH splits are patient-disjoint. Intra-patient splits let rare-class beats from one patient appear on both sides, which would hide exactly the generalization failure we care about. A secondary arm uses PTB-XL with added baseline wander to test the coarse-scale side of the tension: whether denoising flattens the ST changes of myocardial infarction.
Baselines
The naive reference point is no denoising at all. A 1D ResNet classifier is trained directly on noisy beats, with the same Noise Stress Test Database noise used as training augmentation and class-balanced sampling. It has the same parameter budget as the denoiser plus classifier pipeline. Any denoising method that does not beat it on minority recall at low SNR has not earned its place as a preprocessing step. We treat this baseline as the most serious threat on the classification metric.
The classical reference is a bandpass filter plus wavelet thresholding, as used inside the LDTF embedding.
The strongest existing denoiser is the fully convolutional autoencoder of Noise Reduction in ECG Signals Using Fully Convolutional Denoising Autoencoders. It reports 15.49 dB SNR improvement at −1 dB input SNR, against 14.14 dB for a plain CNN and 11.75 dB for a dense network. We reimplement it on our splits and first confirm we land near that figure before trusting our comparisons.
For classification, the strongest reported figures in the file are 98.39% average recall for LDTF with SMOTE and 0.979 AP for TolerantECG. Neither comes from a patient-disjoint, noise-swept protocol with per-class reporting. We therefore rerun LDTF-style features with SMOTE on our protocol rather than compare to the quoted numbers.
The key ablation is the fixed-resolution autoencoder with the same class-balanced weighting applied at all scales. A second ablation is the progressive model with weighting at all scales instead of only the finest $k$. A third is the progressive model without weighting. Together these separate three possible sources of gain: the decomposition, the balancing, and where the balancing is applied.
All models receive the same training data, noise mixtures, epochs-equivalent compute and a 20-trial hyperparameter search on a patient-disjoint validation split. Winning means the following: S- and F-class recall at 0 dB and below improves over the class-balanced fixed autoencoder by more than the seed-to-seed spread across five seeds, and overall SNR improvement stays within 0.5 dB of the unweighted autoencoder.
Why not just…?
Why not just skip denoising and train a noise-augmented classifier end to end? For classification accuracy alone, this may well win. TolerantECG's results suggest robust end-to-end models are strong. The experiment compares directly against it. If it matches or beats every denoise-then-classify pipeline on minority recall, we will say so. The proposal's value then narrows to the cases where a clean waveform is itself the product: displayed telemetry, interval measurement, and clinician review. In those cases the audit of what denoisers erase still matters.
Why not just class-balance the existing fixed-resolution autoencoder? This is the most plausible competitor to the fix, and it is our central ablation. Our reason to expect it to fall short is that balancing at all scales also upweights coarse-scale reconstruction of rare beats. Coarse-scale content is shared across classes and is not where the damage is, so that weighting spends capacity and can leave more fine-band noise behind. The comparison at matched SNR improvement settles the question. If the two match, the progressive structure survives only as an interpretability instrument, and we will report it that way.
Why not just use Jacobian saliency or frequency-band occlusion on a standard denoiser to see what was removed? Jacobian saliency from Clinically meaningful interpretability of an AI model for ECG classification explains a scalar class score. It does not explain a removed waveform. Band occlusion feeds the network out-of-distribution inputs. We compare our scale-swap counterfactual flip rates against occlusion-based band attributions on the same classifier. If the two agree closely, the post-hoc route is adequate, and our decomposition's interpretability advantage is weaker than claimed.
Why not just apply SMOTE after a fixed denoising embedding, as LDTF does? SMOTE interpolates between beats that have already passed through the denoiser. If the audit shows that minority morphology is destroyed before that point, oversampling the damaged representation cannot restore it. The LDTF-with-SMOTE rerun under patient-disjoint splits tests this directly.
Decisive Checks
The first check decides whether the project's premise holds. Do fixed-resolution autoencoders show a per-class fidelity gap, with higher waveform error and more removed QRS-window energy for V, F and S beats than for N beats at the same input SNR? Is that gap concentrated in the finest bands? If not, the tension the research file describes is not real at MIT-BIH scale, and we publish the negative audit.
The second check is whether the scale-swap counterfactuals locate S-class evidence mainly at coarse scales, where prematurity is visible in timing, and V and F evidence at fine scales. This would confirm that the decomposition tracks physiology rather than an arbitrary split.
The third check is the head-to-head at 0 and −6 dB among four models: the weighted progressive denoiser, the class-balanced fixed autoencoder, and the noise-augmented end-to-end classifier, together with the unweighted fixed autoencoder. On PTB-XL with baseline wander, the analogous check is whether coarse-scale denoising preserves ST-segment amplitude for infarction records. The finding that rare beats are erased only at fine scales, while infarction signs are erased only at coarse scales, would itself be a useful result for anyone deploying ECG denoisers on wearables.
References
- Noise Reduction in ECG Signals Using Fully Convolutional Denoising Autoencoders
- Low-Dimensional Denoising Embedding Transformer for ECG Classification
- Progressive Growing of GANs for Improved Quality, Stability, and Variation
- Improving trust and confidence in medical skin lesion diagnosis through explainable deep learning
- Clinically meaningful interpretability of an AI model for ECG classification
- Climbing the Ladder of Interpretability with Counterfactual Concept Bottleneck Models
- TolerantECG: A Foundation Model for Imperfect Electrocardiogram
- Self-supervised representation learning from 12-lead ECG data