Learning Representations Directly from Raw Physiological Waveforms

190 concepts184 papers2 proposals

For decades, analyzing physiological signals like electrocardiograms (ECGs) and brainwaves (EEGs) followed a predictable recipe: an expert clinician or engineer designed a mathematical formula to extract specific metrics—such as heart rate variability, power spectral density, or windowed signal averages—and fed these metrics into a shallow classifier.

While highly interpretable, this traditional paradigm contains a fundamental flaw. As Yun et al. describe it, manual feature engineering acts as a "low-resolution filter." By forcing complex, high-dimensional biological waveforms into a few human-defined boxes, we inadvertently discard subtle, non-linear, and serendipitous patterns that may hold the key to accurate diagnosis and risk stratification.

To overcome this bottleneck, a major research direction has emerged: end-to-end representation learning directly from raw physiological signals. By bypassing manual feature extraction entirely, deep learning models can discover optimal, hierarchical representations of biological data directly from the raw waveform.

The Information Bottleneck of Expert Features

Why do hand-crafted features fall short? The core limitation is that they cap a model's performance at the current boundaries of human domain knowledge. As Rouhi notes, relying exclusively on hand-crafted features limits a system to known physiological markers, creating a bottleneck where the model cannot discover latent patterns that have not yet been formally defined by science.

Consider how this manifests across different clinical domains:

  • Missing Subtle Morphology: In cardiac analysis, traditional feature engineering often focuses on prominent landmarks like the QRS complex. However, Ott et al. demonstrated that by analyzing raw ECG signals directly, an explainable AI framework could identify critical, easily overlooked aging markers in the P-wave that standard metrics filtered out.
  • Ignoring Non-Linear Synergies: In biomechanics, human movement is driven by complex, antagonistic muscle interactions. Hajian et al. found that traditional mathematical features struggle to capture the non-linear spatial and temporal synergies present in high-density electromyography (EMG) sensor data during dynamic movements.
  • Stripping Temporal Dynamics: Biological systems are inherently sequential. Chang et al. point out that hand-engineered features fail to capture the complex, sequential dependencies and transitions inherent in physiological states, such as the boundaries between sleep stages.
Traditional Pipeline:
Raw Signal ──> [ Human Expert Heuristics ] ──> Low-Dim Features ──> Classifier ──> Prediction
                      (Information Loss)

End-to-End Pipeline:
Raw Signal ──────────────────────────────────> [ Deep Neural Network ] ───────────> Prediction
                                                 (Learns Optimal Features)

How End-to-End Learning Preserves Information

End-to-end deep learning solves this problem by combining feature extraction and classification into a single, continuous optimization problem. Instead of a human deciding which frequency bands or peaks matter, the neural network uses layers of learnable filters to discover which aspects of the raw waveform are predictive.

Capturing Local and Global Context

To process raw waveforms effectively, models must capture both instantaneous shape (morphology) and long-term trends. For instance, Miller et al. utilized a 1D convolutional-deconvolutional architecture with skip connections to automatically identify sleep arousals from raw physiological waveforms. The early convolutional layers act as automated frequency and shape detectors, while deeper layers capture the broader temporal context.

For complex tasks like epileptic seizure prediction, the architecture must scale across both time and space. Xu et al. addressed this by designing a hybrid network that applies 1D temporal kernels directly to raw EEG data to capture fast electrical transitions, followed by 2D spatial-temporal kernels to map how those signals propagate across different regions of the brain.

Disentangling Style from Content

Raw physiological signals are notoriously noisy and highly variable between individuals. Representation learning allows models to separate the underlying clinical condition from patient-specific artifacts. Yun et al. developed REGLE, an unsupervised framework that leverages Variational Autoencoders (VAEs) to compress raw clinical waveforms into low-dimensional embeddings. By training on the raw data without human bias, the embeddings captured subtle biological variations that significantly improved the accuracy of downstream polygenic risk scores for diseases like asthma and COPD.

Supervised vs. Unsupervised Representations

When moving away from hand-crafted features, researchers face a choice: should the raw signal representations be learned completely unguided (unsupervised) or driven by diagnostic labels (supervised)?

Sun et al. explicitly investigated this trade-off when predicting individual survival distributions from ECG data. They discovered that features extracted via supervised learning—where a model was pre-trained to recognize over 1,400 distinct medical diagnoses from raw waveforms—significantly outperformed both unsupervised embeddings and traditional clinical measurements.

The intuition behind this success is that supervised pre-training forces the network to build internal representations of known pathological structures (such as specific arrhythmias or blocks). When the model is later tasked with a complex problem like mortality prediction, it can leverage these high-level, learned diagnostic concepts rather than trying to parse the raw, noisy waveform from scratch.

Broad Impact Across Physiological Domains

The transition from expert-defined features to raw signal representations has yielded substantial performance gains across a wide variety of applications:

  • Cardiovascular Risk: Traditional risk scores rely on rigid, rule-based interpretations of ECGs. By applying Multiple Instance Learning directly to raw, long-term ECG recordings, Shanmugam et al. built a model that automatically flags "worrisome" heartbeat segments, significantly outperforming traditional clinical risk metrics.
  • Neurological Monitoring: Traditional seizure detection relies on manual calculations of spectral power or entropy. End-to-end deep learning models processing raw iEEG data learn to identify the subtle, high-dimensional spatial-temporal precursors of a seizure long before they manifest in a way that is obvious to manual inspection (Eberlein et al.).
  • Prosthetics and Robotics: In assistive technologies, mechanical lag can make exoskeletons feel unnatural. Yi demonstrated that processing raw, multi-modal EMG and motion signals through a deep network allows the system to decode non-linear muscle synergies in real time, predicting lower-limb joint angles up to 100ms into the future to eliminate robotic latency.

By treating physiological signals as rich, high-dimensional landscapes rather than a collection of simple statistics, representation learning bridges the gap between raw biological reality and clinical utility.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 178 papers
  2. Version 1 · 2026-10-04 · 141 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_manual_feature_engineering_limitations_feb9b71d,
  title = {manual feature engineering limitations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/manual-feature-engineering-limitations/rdr_feb9b71dad534ee08bcdac72d9c87b98},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_manual_feature_engineering_limitations_feb9b71d_v2,
  title = {manual feature engineering limitations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/manual-feature-engineering-limitations/rdr_feb9b71dad534ee08bcdac72d9c87b98/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake