Source-linked AI summary
A review of feature extraction and performance evaluation in epileptic seizure detection using EEG
Poomipat Boonyakitanont, Apiwat Lek-uthai, Krisnachai Chomtho, Jitkomut Songsiri
TL;DR
Manual seizure detection in continuous EEG is time-consuming, while jointly used features may be redundant and prior evaluations may be inadequate. This paper reviews EEG seizure-detection features and metrics, evaluates individual-feature significance with Bayesian error and probability-distribution estimation, and analyzes redundancy with correlation-based selection. Variance, energy, nonlinear energy, and Shannon entropy from raw EEG, plus selected wavelet-coefficient features, significantly captured seizures and improved Bayesian error from baseline by 4.77–13.51%.
Problem
Manual continuous-EEG seizure review is time-consuming, while prior studies may combine redundant features and lack conclusive evidence about individual feature contributions and fair evaluation.
Method
The paper reviews EEG seizure features and performance metrics, tests individual-feature significance using Bayesian error and non-parametric probability distributions, and analyzes redundancy with correlation-based feature selection.
Results
Variance, energy, nonlinear energy, and Shannon entropy from raw EEG, plus variance, energy, kurtosis, and line length from wavelet coefficients, significantly captured seizures.
Takeaways & Limitations
Feature selection should consider individual significance and redundancy alongside seizure-detection performance metrics.
Takeaways & Limitations
Event-based metrics require care because high detection rates and low false-positive rates can still mislead conclusions about correct detection.
Abstract
from arXiv · showhide
Since the manual detection of electrographic seizures in continuous electroencephalogram (EEG) monitoring is very time-consuming and requires a trained expert, attempts to develop automatic seizure detection are diverse and ongoing. Machine learning approaches are intensely being applied to this problem due to their ability to classify seizure conditions from a large amount of data, and provide pre-screened results for neurologists. Several features, data transformations, and classifiers have been explored to analyze and classify seizures via EEG signals. In the literature, some jointly-applied features used in the classification may have shared similar contributions, making them redundant in the learning process. Therefore, this paper aims to comprehensively summarize feature descriptions and their interpretations in characterizing epileptic seizures using EEG signals, as well as to review classification performance metrics. To provide meaningful information of feature selection, we conducted an experiment to examine the quality of each feature independently. The Bayesian error and non-parametric probability distribution estimation were employed to determine the significance of the individual features. Moreover, a redundancy analysis using a correlation-based feature selection was applied. The results showed that the following features --variance, energy, nonlinear energy, and Shannon entropy computed on a raw EEG signal, as well as variance, energy, kurtosis, and line length calculated on wavelet coefficients-- were able to significantly capture the seizures. An improvement of 4.77--13.51% in the Bayesian error from the baseline was obtained.
1 Introduction
Automated EEG seizure detection research uses diverse features, transformations, classifiers, and evaluation approaches, but prior work leaves feature contributions, redundancy, and fair comparison insufficiently resolved.
- Motivation: Manual review of continuous EEG recordings is time-consuming and requires trained neurologists, motivating automated seizure detection methods.The paper frames automated analysis as a pipeline involving acquisition, preprocessing, feature extraction or channel selection, and classification.
- Prior approaches: Prior studies have applied features from time, frequency, and time-frequency domains, sometimes using large feature groups with machine-learning classifiers.Reviewed approaches include SVMs, ANNs, deep learning, and wavelet-based transformations.
- Prior approaches: 89.2% detection rate with one false detection per hour was reported for 55 features combined with an SVM and post-processing in neonatal seizure detection.This result is an example of prior feature-based seizure detection performance.
- Gaps in prior work: Earlier reviews did not consistently explain wavelet benefits, wavelet-transform mathematics, or the mathematical details of all seizure-analysis processes.These omissions limit the completeness of feature and method interpretation in the reviewed literature.
- Gaps in prior work: Previous studies often combined features without identifying individual contributions, inconsistently stated complicated feature definitions, and relied heavily on accuracy despite imbalanced datasets.The paper argues that selecting distinguished features and using fairer performance comparisons require further review.
- Review scope: The review primarily gathered publications from Google Scholar, ScienceDirect, and IEEE Xplore, focusing mainly on studies published between January 2010 and December 2018.Its organization covers feature descriptions and computational complexity, evaluation metrics, and feature-based AESD methods.
2 Feature extraction
The paper organizes EEG seizure-detection features by signal domain and explains their meanings, computations, and computational considerations. Features quantify distributional, amplitude, complexity, self-similarity, and spectral characteristics of raw or transformed EEG signals.
- Feature domains: Time-domain features are calculated from raw or time-domain preprocessed EEG, frequency-domain features from Fourier transforms, and time-frequency features from STFT or wavelet representations.The feature input sequence may be a raw segment, power spectral density, wavelet coefficients, or intrinsic mode functions.
- Time-domain features: Statistical features such as mean, variance, skewness, and kurtosis discriminate ictal from normal patterns under differing EEG statistical distributions.Minimum and maximum values additionally quantify signal range or magnitude.
- Amplitude features: Energy is the sum of squared signal values, average power is the mean square, and RMS is the square root of average power.These measures are mutually relevant indicators of signal amplitude.
- Amplitude features: Line length measures the total vertical length of a signal and is also called curve length.It is used as a signal-pattern feature in seizure detection literature.
- Amplitude features: Nonlinear energy extends quadratic energy using shifted and lagged sequences, with high values indicating high-frequency oscillation and amplitude for simple harmonic motion.The stated relation is NE proportional to A^2ω^2 when sampling frequency is high.
- Entropy features: Shannon entropy reflects uncertainty in a random process, while approximate, sample, permutation, weighted-permutation, fuzzy, and distribution entropy quantify regularity, complexity, uncertainty, or self-similarity in different ways.Approximate entropy compares similarity patterns, permutation entropy measures local complexity, and distribution entropy uses a histogram of template-vector distances.
- Entropy and nonlinear features: SVDEn is formed from normalized singular values of a matrix built temporally with delayed signal rows or spatially with channel sequences; low SVDEn indicates seizure patterns in noisy EEG.The interpretation follows the observation that singular values associated with noisy signals are smaller than those of deterministic signals.
- Frequency-domain features: Intensity weighted bandwidth measures normalized-PSD width, and it is smaller during seizure activities because seizure PSDs are sharper.The normalized PSD values are denoted x[k].
3 Epileptic seizure detection
This section reviews EEG seizure-detection features, classifiers, databases, and performance metrics, emphasizing feature combinations and evaluation boundaries. It reports that feature selection and classifier choice affect detection results, while epoch- and event-based metrics can misrepresent performance in different ways.
- 3.1 Performance metrics: Performance assessment uses TP, FP, FN, and TN to derive sensitivity, specificity, accuracy, TPR, or FPR.
- 3.1 Performance metrics: Epoch-based metrics aggregate correctly and incorrectly classified EEG segments, but can remain high when a short seizure event is missed.
- 3.1 Performance metrics: Event-based metrics include good detection rate and false positives per hour, yet may count detection at only the final second of a seizure as successful.
- 3.2 Selected features used in epileptic seizure detection: Wavelet-based features achieved 96.87% average accuracy, 72.99% sensitivity, and 98.13% specificity with SVM, compared with 95.17%, 66.35%, and 96.91% for TDFs.
- 3.2.3 Time-frequency-domain features (TFDFs): Joint TFDFs with dimensionality reduction and RBF SVM detected seizures with 98-100% accuracy, while ELM combinations were reported as accurate and quick.
- 3.2.4 Multi-domain features: Across reviewed studies, raw EEG or transformed coefficients were commonly paired with deep learning, whereas statistical parameters were not used individually.
4 Methods for feature evaluation
The paper evaluates individual EEG features with Bayesian error estimation and addresses feature redundancy using correlation-based feature selection. Bayesian probabilities are estimated non-parametrically, while CFS favors features correlated with the class but weakly correlated with one another.
- Bayes error rate: Bayesian error measures how effectively each individual feature separates normal and seizure classes.The method uses posterior probabilities and seeks the minimum classification error achievable by selecting the most probable class.
- Bayes error rate: Non-parametric kernel estimation approximates unknown likelihood distributions and supports posterior-probability calculation through Bayes’ rule.The Gaussian kernel uses class samples, sample size, and a bandwidth based on the sample standard deviation.
- Bayes error rate: Improvement rate compares feature-based Bayesian error with a standard-condition error because conventional metrics do not reveal individual-feature significance.The standard condition is denoted err0, and the improvement rate is proposed for evaluating individual features.
- Correlation-based feature selection (CFS): Correlation-based feature selection evaluates feature subsets using class correlation and feature-feature correlation.CFS assumes a useful subset is strongly related to the class while avoiding redundancy among its features.
5 Experimental results
The experiments evaluated individual EEG features for seizure significance and then analyzed redundancy among significant features. Wavelet-coefficient features, especially energy and variance, produced strong improvement rates and an optimal five-feature subset.
- Experimental setup: The experiment used CHB-MIT Scalp EEG recordings, selecting records containing at least one seizure activity for feature evaluation.The database contains 24 recordings from 23 patients, and two records were randomly chosen from each case subject to the inclusion condition.
- Experimental setup: EEG epochs were formed with 4-second windows advanced by 1 second, and features were computed independently for each epoch and channel.The analysis used commonly employed features and spatially averaged features over the left and right sides of the brain.
- Feature significance: 13.51% was the highest improvement rate, achieved by energy from level D1 of the left-side wavelet coefficients.Wavelet-coefficient variance and energy were relatively strong across decomposition levels; line length and kurtosis were also significant at selected levels.
- Feature redundancy analysis: Five features maximized the CFS merit score: EnergyD1R, VarianceD5L, VarianceD1L, EnergyD5L, and VarianceD1R.The merit score peaked at a subset size of five and decreased as more features were added.
- Feature redundancy analysis: The selected five features also had relatively high improvement rates and required O(N) total computation, supporting their use in AESD.The authors report that classification accuracy did not increase when additional features were added to the optimal subset.
6 Conclusions
The paper reviews EEG seizure-detection features, their mathematical descriptions, interpretations, and classification uses, then tests feature significance and redundancy. It finds raw-signal variance, energy, nonlinear energy, and Shannon entropy useful, while wavelet-coefficient energy and variance are recommended because of significance and low computational complexity.
- Review contribution: The review provides detailed mathematical descriptions, computations, usages, and intuitive meanings for EEG features used in epileptic seizure detection.
- Review findings: Energy was the feature most widely used to quantify a relation with signal amplitude in the reviewed literature.
- Experimental conclusions: Raw-signal variance, energy, nonlinear energy, and Shannon entropy improved discrimination between seizure and normal EEG signals relative to the baseline.Other time-domain and frequency-domain features produced insignificant seizure-detection outcomes in the experiment.
- Experimental conclusions: Wavelet-coefficient energy and variance from selected decomposition levels were recommended for online and offline AESD because of their significance and low computational complexity.