Source-linked AI summary
Energy Distribution of EEG Signals: EEG Signal Wavelet-Neural Network Classifier
Ibrahim Omerhodzic, Samir Avdakovic, Amir Nuhanovic, Kemal Dizdarevic
TL;DR
The paper addresses limited EEG classification and the need to identify epileptic activity by proposing a wavelet-based neural-network classifier. It extracts multiresolution energy-distribution features with DWT and MRA, then classifies them with a neural network; across 300 EEG signals, the reported accuracy was 94.0%.
Problem
EEG classification and evaluation are limited, while visual analysis may be insufficient for routine diagnosis of epileptic activity.
Method
The method decomposes EEG signals with DWT and MRA, extracts percentage energy features using Parseval’s theorem, and classifies those features with a neural network.
Results
94.0% accuracy was reported, with 47 of 50 testing EEG signals correctly classified and 100% correct classification for normal EEG signals.
Takeaways & Limitations
Percentage energy distributions can be used for EEG-signal classification, including distinguishing healthy, epilepsy, and seizure-related signals.
Abstract
from arXiv · showhide
In this paper, a wavelet-based neural network (WNN) classifier for recognizing EEG signals is implemented and tested under three sets EEG signals (healthy subjects, patients with epilepsy and patients with epileptic syndrome during the seizure). First, the Discrete Wavelet Transform (DWT) with the Multi-Resolution Analysis (MRA) is applied to decompose EEG signal at resolution levels of the components of the EEG signal (delta, theta, alpha, beta and gamma) and the Parsevals theorem are employed to extract the percentage distribution of energy features of the EEG signal at different resolution levels. Second, the neural network (NN) classifies these extracted features to identify the EEGs type according to the percentage distribution of energy features. The performance of the proposed algorithm has been evaluated using in total 300 EEG signals. The results showed that the proposed classifier has the ability of recognizing and classifying EEG signals efficiently.
I. INTRODUCTION
The paper motivates automated EEG classification because epilepsy is widespread, seizure detection matters clinically, and manual or visual EEG analysis is limited. It proposes combining wavelet-based energy features with neural-network classification.
- More than 40 million people worldwide are affected by epilepsy, whose seizures can have debilitating medical and social consequences.
- EEG is important for identifying epileptic seizure activity, but expert-based manual analysis is limited by the small number of skilled professionals.
- Visual time-domain EEG analysis may be insufficient because classification and evaluation are limited and experts lack a definite criterion.
- The paper positions wavelet-transform and neural-network classification as an established approach previously applied to EEG and other medical signal-processing problems.
- The proposed algorithm decomposes EEG signals with DWT and MRA, extracts percentage energy distributions, and classifies them with a neural network.
A. Wavelet Transform
The wavelet-transform section develops a time-frequency representation and uses multiresolution decomposition to separate signal approximations and details. Parseval’s theorem then partitions signal energy across decomposition levels.
- The wavelet transform represents a function in the time-frequency domain using a zero-average wavelet function.
- The CWT uses a mother wavelet, complex conjugation, and scaling and translation parameters to analyze signal behavior across frequency and position.
- The DWT applies discrete scaling and translation values, making it the wavelet transform typically used in engineering applications.
- Choosing a dyadic-orthonormal wavelet with a₀=2 and b₀=1 provides the basis for multi-resolution analysis.
- In MRA, a time series is decomposed into low-pass approximations from scaling functions and high-pass details from wavelets.
- Parseval’s theorem partitions signal energy across decomposition levels, including detail energies at each level and approximation energy at the final level.
B. Artificial Neural Networks
The paper describes artificial neural networks as weighted neuron-like systems trained to map feature vectors to outputs. It uses backpropagation for classification, leveraging neural networks’ pattern-recognition capabilities.
- Artificial neural networks simulate low-level biological-neuron functions, with knowledge distributed across neurons and connection weights.
- During training, feature vectors are supplied as inputs while weights and biases are adjusted to capture relationships between patterns and outputs.
- The network uses backpropagation, identified as the most frequently used training algorithm for classification problems in this work.
- Artificial neural networks are useful for complex pattern-recognition and classification tasks and are widely applied in biomedical modeling and diagnostic classification.
- Neural-network architectures differ in how they learn, with their details documented across the literature.
C. Proposed Methodology
The methodology decomposes EEG signals into clinically relevant frequency components, converts their wavelet coefficients into energy-distribution features, and classifies those features with a neural network.
- Data and signal representation: The study uses EEG datasets from healthy subjects, patients with epilepsy syndrome before seizures, and patients recorded during seizures.The datasets comprise Bonn EEG groups A, C, and E, with 23.6-second single-channel segments sampled at 173.61 Hz.
- Wavelet decomposition: DWT with MRA decomposes EEG signals into detail and approximation components aligned with gamma, beta, alpha, theta, and delta bands.Daubechies 4 is selected, and the signals are decomposed into D1-D5 and final approximation A5.
- Classification pipeline: The classification pipeline transforms the preprocessed EEG from the time domain into the wavelet domain before neural-network processing.Feature extraction reduces the original waveform to a lower-dimensional vector containing useful information for pattern recognition.
- Neural-network classifier: A feed-forward neural network receives the wavelet energy-distribution features and learns input-output relationships for EEG classification.The network is trained by adjusting connection weights and biases.
CLASSIFICATION RESULTS
The classifier distinguishes three EEG groups using six wavelet-energy features. Their energy distributions differ across groups, and the trained WNN achieves strong test-set classification performance.
- Experimental setup: The experiment analyzes 300 EEG signals from healthy, epilepsy-syndrome steady-state, and seizure groups using percentage energy distributions.Each group contributes 100 signals, represented through six features: D1, D2, D3, D4, D5, and A5.
- Energy-distribution findings: Healthy EEGs show approximately 20% energy in both beta and alpha components, about 10% in theta, and approximately 5% in gamma.Noise is described as negligibly small, while the D5 component is about 45% for the healthy signals overall, with variation across samples.
- Energy-distribution findings: Epilepsy-syndrome EEGs have lower D2-D4 energy and larger D5 and A5 energy than healthy EEGs.The seizure-group distribution is distinct, with D3, D4, and D5 dominant and A5 somewhat lower.
- Classification performance: The percentage distribution of EEG energy can be used as features for classifying EEG signals.The results support using wavelet-energy distributions as inputs to the neural-network classifier.
- Experimental setup: The WNN is trained with 250 datasets and tested on 50 datasets after selecting its architecture and training parameters through experiments.The tested parameters include hidden-layer size, momentum constant, learning rate, and activation functions.
- Classification performance: 94.0% accuracy was achieved on the testing set, with 47 of 50 EEG signals correctly classified.Normal EEG signals achieved a 100% correct-classification rate.
IV. CONCLUSION
The proposed WT-based classifier uses multiresolution energy features and a feedforward neural network to classify EEG signals efficiently. It also reduces data and computational requirements during classification.
- DWT with MRA decomposes EEG signals into resolution levels corresponding to delta, theta, alpha, beta, and gamma components.
- Parseval’s theorem extracts the percentage distribution of EEG energy features across resolution levels.
- A feedforward neural network classifies the extracted energy features to identify EEG types.
- The classifier recognizes and classifies EEG signals efficiently.
- The method reduces data size and indicates the main characteristics of the signal.It can also reduce memory space, preprocessing needs, and network size while increasing computation speed.