Deep Learning for Raw Physiological Signal Analysis

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2141 papers142 conceptsfrozen 2026-10-04

For decades, the analysis of physiological signals—the electrical rhythms of the heart (ECG), the firing of neurons (EEG), or the tension in muscles (sEMG)—relied on a process called manual feature engineering. Experts would define specific mathematical descriptors, such as the height of a wave or the power of a frequency band, to summarize complex biological data.

However, a massive body of research is now shifting toward a different paradigm: end-to-end deep learning. Instead of telling a model what to look for, researchers are feeding raw, unsegmented waveforms into neural networks, allowing the machines to discover "hidden" biomarkers that human eyes and traditional statistics often miss.

The Limitation of Hand-Crafted Features

Traditional diagnostic systems are often "lossy." When a human expert defines a feature—like the ST-segment elevation in an ECG—they are essentially applying a filter that discards any information not fitting that definition. As Nils Strodthoff and Wilhelm Haverkamp argue in EKG-Diagnostik mithilfe künstlicher Intelligenz, these rigid, rule-based approaches fail to account for the non-linear complexity of biological systems.

Manual features struggle with:

  • Subjectivity: Different experts may interpret the same signal differently.
  • Noise: In real-world settings like the ICU, movement and technical artifacts often break rule-based systems, leading to "alarm fatigue" Sajad Mousavi.
  • Invisibility: Humans are limited by visual perception. We cannot easily see the subtle, high-dimensional correlations that might predict a heart attack hours before it happens.

Learning the Language of the Body

The core intuition behind the new approach is to treat physiological signals as a language. In Sleep Stage Classification Based on Multi-level Feature Learning, Eric I-Chao Chang and colleagues treat heart rate and movement data as a "vocabulary" of sleep patterns. By using Bidirectional Long Short-Term Memory (BLSTM) networks, the model learns the "grammar" of how one physiological state transitions into another.

This shift is visible across diverse domains:

  • Neurology: Yankun Xu developed an end-to-end model for seizure prediction that uses raw EEG signals, bypassing the need for experts to hand-select frequency bands.
  • Cardiology: The PreOpNet model, developed by John Theurer and Bryan He, analyzes raw 12-lead ECG waveforms to predict post-operative mortality, outperforming traditional clinical risk scores.
  • Physical Therapy: Nana Wang created EasiCSDeep, which reorganizes raw sEMG signals into anatomically-inspired structures for a CNN to identify cervical spondylosis.

Architectural Strategies for Waveforms

How do these models actually "see" the signal? Researchers generally use two main strategies to bridge the gap between raw data and diagnostic insight.

1. 1D Convolutional Neural Networks (1D-CNNs)

Since physiological signals are temporal, 1D-CNNs are used to slide filters across the time axis. This allows the model to detect local morphological patterns—like a specific "spike" or "dip"—regardless of when they occur. Imanne El Maâchi utilized 18 parallel 1D-CNNs to detect Parkinson’s Disease from gait patterns, capturing the complex, non-linear noise of sensor data.

2. Time-Frequency Transformations

Sometimes, the most important information is hidden in how frequencies change over time. Models like the FE-CNN by Qingsong Xie transform 1D signals into 2D spectrograms. This "feature enrichment" allows standard image-recognition CNNs to identify structural patterns in the signal's frequency heartbeats.

Beyond Diagnosis to Prediction

The most provocative result of this research direction is the discovery of "invisible" biomarkers. In An ECG-based artificial intelligence model for assessment of sudden cardiac death risk, Lauri Holmström demonstrated that AI can identify patients at risk of sudden death even when their ECG looks "normal" to a cardiologist.

Similarly, Antonio H. Ribeiro found that the internal features learned by these networks correlate with specific genetic risk factors and long-term mortality. By moving away from human-defined heuristics, we are not just automating the doctor; we are uncovering a deeper layer of biological information that was previously inaccessible.

Papers in this version

Showing 60 of 141 papers.

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 1, frozen text):

@misc{lacuna_deep_learning_for_raw_physiological_sign_feb9b71d_v1,
  title = {Deep Learning for Raw Physiological Signal Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-learning-for-raw-physiological-signal-analysis/rdr_feb9b71dad534ee08bcdac72d9c87b98/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}