Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WaveNet's Precision in EEG Classification

Casper van Laar, Khubaib Ahmed

TL;DR

This work demonstrates that a WaveNet architecture, originally designed for raw audio, can effectively classify EEG signals into physiological, pathological, artifact, and noise categories. By leveraging dilated causal convolutions and residual/skip connections, the model captures both fine-grained and long-range temporal dependencies, and is trained with adaptive dropout and focal loss to handle class imbalance. The approach achieves a high average F1 score and strong noise/artifact discrimination, while remaining robust across data from multiple hospitals, albeit with some confusion between physiological and pathological patterns. The study shows practical potential for scalable, real-time EEG analysis and suggests avenues for cross-domain extensions and interpretability, including applications to spike sorting and integration with hybrid architectures.

Abstract

This study introduces a WaveNet-based deep learning model designed to automate the classification of EEG signals into physiological, pathological, artifact, and noise categories. Traditional methods for EEG signal classification, which rely on expert visual review, are becoming increasingly impractical due to the growing complexity and volume of EEG recordings. Leveraging a publicly available annotated dataset from Mayo Clinic and St. Anne's University Hospital, the WaveNet model was trained, validated, and tested on 209,232 samples with a 70/20/10 percent split. The model achieved a classification accuracy exceeding previous CNN and LSTM-based approaches, and was benchmarked against a Temporal Convolutional Network (TCN) baseline. Notably, the model distinguishes noise and artifacts with high precision, although it reveals a modest but explainable degree of misclassification between physiological and pathological signals, reflecting inherent clinical overlap. WaveNet's architecture, originally developed for raw audio synthesis, is well suited for EEG data due to its use of dilated causal convolutions and residual connections, enabling it to capture both fine-grained and long-range temporal dependencies. The research also details the preprocessing pipeline, including dynamic dataset partitioning and normalization steps that support model generalization.

WaveNet's Precision in EEG Classification

TL;DR

This work demonstrates that a WaveNet architecture, originally designed for raw audio, can effectively classify EEG signals into physiological, pathological, artifact, and noise categories. By leveraging dilated causal convolutions and residual/skip connections, the model captures both fine-grained and long-range temporal dependencies, and is trained with adaptive dropout and focal loss to handle class imbalance. The approach achieves a high average F1 score and strong noise/artifact discrimination, while remaining robust across data from multiple hospitals, albeit with some confusion between physiological and pathological patterns. The study shows practical potential for scalable, real-time EEG analysis and suggests avenues for cross-domain extensions and interpretability, including applications to spike sorting and integration with hybrid architectures.

Abstract

This study introduces a WaveNet-based deep learning model designed to automate the classification of EEG signals into physiological, pathological, artifact, and noise categories. Traditional methods for EEG signal classification, which rely on expert visual review, are becoming increasingly impractical due to the growing complexity and volume of EEG recordings. Leveraging a publicly available annotated dataset from Mayo Clinic and St. Anne's University Hospital, the WaveNet model was trained, validated, and tested on 209,232 samples with a 70/20/10 percent split. The model achieved a classification accuracy exceeding previous CNN and LSTM-based approaches, and was benchmarked against a Temporal Convolutional Network (TCN) baseline. Notably, the model distinguishes noise and artifacts with high precision, although it reveals a modest but explainable degree of misclassification between physiological and pathological signals, reflecting inherent clinical overlap. WaveNet's architecture, originally developed for raw audio synthesis, is well suited for EEG data due to its use of dilated causal convolutions and residual connections, enabling it to capture both fine-grained and long-range temporal dependencies. The research also details the preprocessing pipeline, including dynamic dataset partitioning and normalization steps that support model generalization.
Paper Structure (11 sections, 5 figures, 3 tables)

This paper contains 11 sections, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Dataset
  4. Methods
  5. Preprocessing steps
  6. Model structure
  7. Baseline: Temporal Convolutional Network (TCN)
  8. Training summary
  9. Results
  10. Conclusion
  11. Detailed Training Configuration

Figures (5)

  • Figure 1: Healthy physiological signals. Characterized by a steep incline and decline before and after the peak.
  • Figure 2: Example of abnormal brain activity observed in epilepsy-prone neurons is identified by a sharp wave transient. Indicated by the red arrow one can see this abnormal transient, a high frequency oscillation (HFO) can be observed riding atop the peak of the spikes.
  • Figure 3: Accuracy figure, plotted every 150 batches each batch contains 32 samples for the training set. Yellow for the validation set plotted every epoch.
  • Figure 4: Validation and training loss figure, plotted every 150 batches each batch contains 64 samples for the training set. Yellow for the validation set plotted every epoch.
  • Figure A5: Dropout rate dynamics throughout training, governed by the composite performance score.