Learning Robust Representations from Noisy Physiological Signals

33 concepts33 papers4 proposals

In an ideal clinical setting, physiological signals like electrocardiograms (ECGs) or electroencephalograms (EEGs) are pristine, easily interpretable waveforms. In the real world—such as intensive care units, ambulatory monitoring, or consumer smartwatches—these signals are continuously degraded by environmental and biological noise. As noted by Lele Wang et al. in Performance Optimization of Support Vector Machine with Adversarial Grasshopper Optimization for Heart Disease Diagnosis and Feature Selection, power line interference, patient movement, and muscle tremors introduce severe distortions that obscure critical diagnostic markers.

The fundamental challenge of physiological denoising is spectral overlap. Traditional signal processing relies on linear bandpass filters to isolate frequencies of interest. However, the high-frequency components of a muscle tremor or the low-frequency components of respiration (baseline wander) directly overlap with the frequencies of essential clinical features, such as the sharp QRS complex of an ECG. If a filter is too aggressive, it rounds off these peaks, destroying the morphology needed to detect conditions like myocardial infarction. If it is too weak, the automated system misinterprets noise spikes as pathological events, triggering false alarms and clinical fatigue.

To solve this problem, modern research focuses on learning robust, noise-invariant representations that separate true biological pathology from environmental artifacts.

Denoising via Latent Space Bottlenecks

To overcome spectral overlap, modern architectures move away from filtering raw waveforms and instead project signals into constrained latent spaces. The intuition is simple: while noise is high-dimensional and stochastic, valid physiological signals follow a highly structured, low-dimensional manifold. By forcing the signal through a narrow geometric bottleneck, a neural network is physically unable to reconstruct the noise.

In Noise Reduction in ECG Signals Using Fully Convolutional Denoising Autoencoders, Hsin-Tien Chiang et al. leverage this principle using a Fully Convolutional Network (FCN) within a denoising autoencoder framework. The network learns to map corrupted inputs to clean target waveforms by encoding the signal into a compressed representation that captures only the core geometric structure of the heartbeat, effectively stripping away baseline wander and high-frequency artifacts.

To prevent deep learning models from inadvertently memorizing noise patterns within their internal attention mechanisms, Jian Guan et al. introduced the Low-dimensional Denoising Embedding Transformer (LDTF) in Low-dimensional Denoising Embedding Transformer for ECG Classification. Instead of feeding high-dimensional convolutional embeddings into a Transformer—which can preserve and amplify subtle noise artifacts—the authors replace them with a highly constrained, 18-dimensional denoising feature set. This mathematical bottleneck ensures that the self-attention layers focus exclusively on long-range pathological rhythms rather than local environmental interference.

Other approaches, such as the one proposed by Xiaoying Tian et al. in Electrocardiogram Signal Denoising Using Extreme-Point Symmetric Mode Decomposition and Nonlocal Means, combine adaptive signal decomposition with non-local smoothing to isolate artifacts while preserving sharp, non-linear diagnostic details.

Bypassing Reconstruction with Invariant Tokenization

While denoising autoencoders work well when clean target data is available for training, pretraining large foundation models often relies on massive, unannotated datasets that are inherently noisy. If a model is trained using a standard masked autoencoder objective (reconstructing the raw signal from a masked input), it will faithfully learn to replicate the environmental noise present in the training data.

To build true noise immunity, researchers use self-supervised learning objectives that prioritize semantic continuity over pixel-perfect reconstruction. In Self-supervised representation learning from 12-lead ECG data, Temesgen Mehari et al. demonstrate that Contrastive Predictive Coding (CPC) allows models to learn robust cardiac representations. Instead of predicting the exact shape of a future waveform, the model learns to maximize the mutual information between the latent representations of adjacent segments, making the learned features inherently robust to localized noise corruptions.

For highly volatile and noisy periodic signals like brainwaves, Haokun GUI et al. developed a vector quantization framework in Vector Quantization Pretraining for EEG Time Series with Random Projection and Phase Alignment. Instead of direct signal reconstruction, the authors use a parameter-free random projection quantizer combined with Fourier-based phase alignment to map continuous, noisy EEG signals into discrete "semantic units" (the equivalent of words in language models).

Because the quantization step maps a continuous range of variations to a single discrete token, minor fluctuations caused by electrode instability or muscle movement are completely swallowed up. The model pretrains on these stable semantic units, ensuring it captures the underlying neurological state rather than replicating technical artifacts.

Explicit Artifact Modeling and Uncertainty Quantification

In many real-world scenarios, noise is so severe that the underlying physiological signal is completely obliterated. In these cases, attempting to denoise the signal is dangerous, as it can lead to algorithmic hallucinations. Robust systems must therefore learn to explicitly identify noise and quantify their own uncertainty.

In SPINDLE: End-to-end learning from EEG/EMG to extrapolate animal sleep scoring across experimental settings, labs and species, Stefan Bauer demonstrates that explicitly modeling signal artifacts as a distinct, valid category allows automated systems to emulate human expert judgment. Rather than trying to force-fit a corrupted epoch into a biological sleep stage, the model categorizes it as an artifact and discards it, preventing noise from contaminating downstream phenotypic analysis.

When data cannot be discarded, such as in continuous intensive care monitoring, models must communicate their confidence to clinicians. In Quantifying deep neural network uncertainty for atrial fibrillation detection with limited labels, Brian Chen et al. address this by quantifying aleatoric uncertainty—the uncertainty driven by technical interference and sensor displacement. By outputting a clear confidence score alongside a diagnosis, the system allows medical staff to instantly distinguish between a true arrhythmia and a loose electrode lead.

This explicit differentiation is equally vital for hardware safety. In Artificial intelligence for detection of ventricular oversensing: Machine learning approaches for noise detection within nonsustained ventricular tachycardia episodes remotely transmitted by pacemakers and implantable cardioverter-defibrillators, Marc Strik et al. train deep learning models to identify "ventricular oversensing"—non-physiological signals caused by lead fractures or electromagnetic interference that mimic tachycardia. By isolating these artifacts, the system prevents inappropriate pacemaker shocks while alerting clinicians to hardware failures.

Finally, rather than relying on static processing chains, systems can dynamically adapt. Guy Carrault et al. in Pilotage d'algorithmes pour un diagnostic médical robuste en cardiologie propose a "pilot" architecture that monitors real-time noise levels and dynamically adjusts its signal processing algorithms, ensuring robust diagnostic performance across changing clinical environments.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_physiological_signal_noise_35a26c11,
  title = {physiological signal noise},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physiological-signal-noise/rdr_35a26c11b4a14209b0b229d2878b0399},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_physiological_signal_noise_35a26c11_v1,
  title = {physiological signal noise},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physiological-signal-noise/rdr_35a26c11b4a14209b0b229d2878b0399/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake