PaperlayerView exact revision on arXiv

Source-linked AI summary

Detection of Epileptic Seizures on EEG Signals Using ANFIS Classifier, Autoencoders and Fuzzy Entropies

Afshin Shoeibi, Navid Ghassemi, Marjane Khodatars, Parisa Moridian, Roohallah Alizadehsani, Assef Zare, Abbas Khosravi, Abdulhamit Subasi, U. Rajendra Acharya, J. Manuel Gorriz

arXiv:2109.04364v2eess.SPcs.LG

TL;DR

EEG seizure detection is challenged by subjective, time-consuming visual review and artifacts. The paper combines TQWT, fuzzy-entropy features, autoencoder dimensionality reduction, and optimized ANFIS classifiers; ANFIS-BS reports high accuracy across Bonn and Freiburg evaluations.

  • Problem

    Visual EEG seizure review is subjective and time-consuming, and myogenic and ocular artifacts make detection challenging.

    Source: §I · p. 1

  • Method

    The approach extracts fuzzy-entropy features from TQWT sub-bands, reduces their dimensionality with an autoencoder, and compares ANFIS classifiers optimized with GOA, PSO, or BS.

    Source: §I · p. 2, B. Preprocessing using TQWT · p. 3

  • Results

    ANFIS-BS achieved 99.74% accuracy for Bonn two-class classification, 99.46% for Bonn three-class classification, and 99.28% on Freiburg.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The authors say the approach could aid physicians and other medical professionals in the future as a software or hardware diagnostic platform.

    Source: §VI · p. 13

    Feature extraction's high computational cost may limit applications, and fuzzy-entropy hyperparameters are usually set by trial and error.

    Source: §V · p. 10

Abstract

from arXiv · show

Epileptic seizures are one of the most crucial neurological disorders, and their early diagnosis will help the clinicians to provide accurate treatment for the patients. The electroencephalogram (EEG) signals are widely used for epileptic seizures detection, which provides specialists with substantial information about the functioning of the brain. In this paper, a novel diagnostic procedure using fuzzy theory and deep learning techniques is introduced. The proposed method is evaluated on the Bonn University dataset with six classification combinations and also on the Freiburg dataset. The tunable-Q wavelet transform (TQWT) is employed to decompose the EEG signals into different sub-bands. In the feature extraction step, 13 different fuzzy entropies are calculated from different sub-bands of TQWT, and their computational complexities are calculated to help researchers choose the best set for various tasks. In the following, an autoencoder (AE) with six layers is employed for dimensionality reduction. Finally, the standard adaptive neuro-fuzzy inference system (ANFIS), and also its variants with grasshopper optimization algorithm (ANFIS-GOA), particle swarm optimization (ANFIS-PSO), and breeding swarm optimization (ANFIS-BS) methods are used for classification. Using our proposed method, ANFIS-BS method has obtained an accuracy of 99.74% in classifying into two classes and an accuracy of 99.46% in ternary classification on the Bonn dataset and 99.28% on the Freiburg dataset, reaching state-of-the-art performances on both of them.

I. INTRODUCTION

The paper proposes an EEG seizure-detection system combining TQWT sub-bands, multiple fuzzy-entropy features, autoencoder dimensionality reduction, and optimized ANFIS classifiers. It evaluates this approach on the Bonn and Freiburg datasets.

I. INTRODUCTION

  • EEG seizure diagnosis by visual inspection is subjective and time-consuming, while artifacts make detection challenging.The introduction describes computer-aided diagnosis systems as supporting automatic seizure identification.Source: §I · p. 1
  • TQWT decomposes Bonn and Freiburg EEG signals into sub-bands, using J = 8, r = 1, and Q = 3.The paper describes TQWT as an improved DWT suitable for preprocessing chaotic biomedical signals such as EEG.Source: §I · pp. 1–2
  • The paper situates fuzzy entropies among nonlinear EEG features and reviews earlier seizure-detection studies using individual fuzzy-entropy variants.Examples include FuEn, FuPeEn, and FuApEn paired with different classifiers or preprocessing approaches.Source: §I · p. 2
  • The study combines 13 fuzzy entropies extracted from TQWT sub-bands, addressing a gap in prior seizure-detection work.The authors state that prior work had not combined various fuzzy entropies as EEG feature extractors for seizure detection.Source: §I · p. 2
  • A hand-tailored autoencoder reduces feature dimensionality by learning a representation in a smaller space.The paper motivates dimensionality reduction by noting that high-dimensional feature vectors can prevent the classifier from training properly.Source: §I · p. 2
  • The classification comparison includes standard ANFIS and GOA-, PSO-, and BS-optimized variants, with BS described as a GA–PSO combination.The authors state that, to their knowledge, this is the first use of BS to train ANFIS.Source: §I · p. 2

II. MATERIAL AND METHODS

The study uses Bonn and Freiburg EEG datasets and applies TQWT preprocessing to decompose signals into frequency sub-bands. The datasets differ in their recording populations and setups, while the TQWT parameters are specified for the method.

1) Bonn Dataset:

  • 500 Bonn signal frames span five classes, with 100 recordings per class, 23.6-second segments, and a 173.61 Hz sampling frequency.Classes A and B use surface EEG from healthy controls; C, D, and E use intracranial recordings from epilepsy patients.Source: 1) Bonn Dataset · p. 3

2) Freiburg Dataset:

  • Freiburg includes intracranial EEG from 21 patients, recorded at 256 Hz using depth, strip, and grid electrodes.The dataset includes ictal, pre-ictal, and inter-ictal samples; patients had two to five observed seizures each.Source: 2) Freiburg Dataset · p. 3

B. Preprocessing using TQWT

  • The paper illustrates TQWT applied to a frame of the Bonn dataset’s signals with r = 3, Q = 1, and J = 8, and shows its frequency response at those settings.TQWT is described as an improved, particular type of DWT used to analyze EEG signals.Source: B. Preprocessing using TQWT · p. 3

C. Feature Extraction

The feature-extraction section applies fuzzy entropies to EEG signals, which the paper characterizes as chaotic. Standard fuzzy entropy forms time-series vectors and compares their pattern similarities using a fuzzy membership function.

C. Feature Extraction

  • Different fuzzy entropies are used to extract features from EEG signals, which the paper describes as having chaotic behavior.The section introduces several entropy methods after stating that entropies are an important class of feature-extraction methods.Source: C. Feature Extraction · p. 4

1) Standard Fuzzy Entropy:

  • Standard FuEn forms length-m vectors from the input time series and subtracts a baseline x0(i) from each vector.The input is written as x(i), for i = 1, 2, ..., N.Source: 1) Standard Fuzzy Entropy · p. 4
  • FuEn measures pattern similarity with a fuzzy membership function, using r and n as the exponential function's predefined gradient and width.The distance d_ij is described as the maximum absolute difference between vectors; the membership function replaces the Heaviside function.Source: 1) Standard Fuzzy Entropy · p. 4

2) Averaged Fuzzy Entropy:

Averaged fuzzy entropy is presented as an improved model of fuzzy entropy. Its construction compares transformed patterns and averages four entropy variants.

2) Averaged Fuzzy Entropy:

  • AFuEn is presented as a novel, improved model of FuEn.The method compares an improved m-pattern with the original pattern.Source: 2) Averaged Fuzzy Entropy · p. 5
  • AFuEn uses four pattern operations: translation, reflection, inversion, and glide reflection.The paper labels these operations T, R, I, and G, respectively.Source: 2) Averaged Fuzzy Entropy · p. 5
  • FuEna is calculated as the average of FuEnT, FuEnR, FuEnI, and FuEnG.These are the four fuzzy entropy variants associated with the defined operations.Source: 2) Averaged Fuzzy Entropy · p. 5

3) Multi-Scale Fuzzy Entropy:

The section describes multi-scale fuzzy entropy and refined composite multiscale fuzzy entropy, then introduces fractional fuzzy entropy based on fractional Shannon entropy. It specifies the scale-series terms and selected parameters for the refined composite method.

3) Multi-Scale Fuzzy Entropy:

  • Multi-scale fuzzy entropy is described as detecting slight variations in EEG signals with high precision.Its scale-series construction uses τ as the scale factor, and the resulting series is used to calculate MFuEn.Source: 3) Multi-Scale Fuzzy Entropy · p. 5
  • For MFuEn, yτ(j) denotes the scale-series structure, τ is the scale factor, ui are independent time-series components, and N is the total number of EEG data points.The paper states that MFuEn is calculated for the resulting yτ(j) series.Source: 3) Multi-Scale Fuzzy Entropy · p. 5

4) Refined Composite Multiscale Fuzzy Entropy:

  • RCMFuEn has µ-based and σ-based techniques, which differ in the equations used in the first algorithm step.The selected tolerance r, fuzzy entropy power n, and embedding dimension m are 0.15, 2, and 2, respectively.Source: 4) Refined Composite Multiscale Fuzzy… · p. 5

5) Fractional Fuzzy Entropy:

  • Fractional fuzzy entropy is introduced using the base idea of fractional Shannon entropy, with fractional information described by order α.The paper cites for the fractional Shannon entropy and its fuzzy-based variant.Source: 5) Fractional Fuzzy Entropy · p. 5

6) Fuzzy Approximate Entropy:

The section describes fuzzy approximate entropy and a modified fuzzy entropy that uses relative energy as its membership function.

6) Fuzzy Approximate Entropy:

  • FuApEn loosens hard boundaries by using a points approach rather than the Heaviside function, with similarity determined by a fuzzy membership function.The similarity calculation uses a predetermined tolerance r.Source: 6) Fuzzy Approximate Entropy · p. 5

7) Minimum Variance Modified Fuzzy Entropy:

  • MoFuEn uses relative energy as its membership function to address limitations of membership functions.The relative energy function ψ is estimated from a time series segmented with length N.Source: 7) Minimum Variance Modified Fuzzy… · p. 6
  • MVMFuEn is defined using Hnew, where H is the set of modified fuzzy entropy values.The supplied text identifies Hnew as a variable but gives no further explanation of its role.Source: 7) Minimum Variance Modified Fuzzy… · p. 6

8) Inherent Fuzzy Entropy:

These subsections describe several fuzzy-entropy procedures, including decomposition-based, distribution-based, cross-signal, and permutation-based calculations.

8) Inherent Fuzzy Entropy:

  • IFuEn decomposes the signal with EMD, evaluates complexity using FuEn, then applies an MFuEn-like stage.The first stage generates IMFs and reconstructs the signal; the final stage follows the MFuEn approach described by relations (13) and (14).Source: 8) Inherent Fuzzy Entropy · p. 6

9) Fuzzy Distribution Entropy:

  • FuDistEn reconstructs state space, forms a Chebyshev-distance matrix, estimates a probability density, then calculates entropy.Its probability estimate excludes diagonal elements to avoid self-matching; the method uses exponential and fuzzy membership functions.Source: 9) Fuzzy Distribution Entropy · p. 6

10) Cross Fuzzy Entropy:

  • CFuEn analyzes similarity or synchrony between two signals using an exponential function instead of the Heaviside function.Its parameters are expressed as the negative logarithm of a conditional probability.Source: 10) Cross Fuzzy Entropy · p. 6

11) Fuzzy Permutation Entropy:

  • FuPeEn orders values from embedded time-series rows, constructs vectors from the reordered series, then applies standard FuEn steps.The embedding uses permuted dimension pm and time dimension τ.Source: 11) Fuzzy Permutation Entropy · p. 6

12) Hierarchical Measure Entropy:

The section characterizes hierarchical and measure fuzzy entropies and compares the computational complexity of the feature calculations.

12) Hierarchical Measure Entropy:

  • HFuEn computes FuEn for each hierarchical component and plots the results against the scale factor.At scale k, uk,e denotes a high- or low-frequency component, and k identifies the hierarchical-analysis layer.Source: 12) Hierarchical Measure Entropy · p. 6

13) Fuzzy Measure Entropy:

  • FuMeEn is described as a generalization of FuEn.The section gives its relation and refers readers to for further information.Source: 13) Fuzzy Measure Entropy · p. 7

D. Computational Complexity of Features

  • FuEn has complexity O(n2m); RCMFuE is O(τn2m), while FuDistEn complexity depends strongly on the density-estimation implementation.Other advanced fuzzy entropies generally have complexity similar to FuEn, though hidden constants can change runtime; MFuEn scales with FuEn.Source: D. Computational Complexity of Features · p. 7

E. Dimensionality reduction using Autoencoder

The section presents an autoencoder for reducing feature dimensionality and describes the standard ANFIS classifier’s architecture.

E. Dimensionality reduction using Autoencoder

  • An autoencoder encodes input features into a smaller latent space and reconstructs them with a decoder, providing dimensionality reduction.The AE is trained with Adadelta and mean square error; its final layer uses Tanh activation, with normalized data matched to that output.Source: E. Dimensionality reduction using… · p. 7
  • The paper presents a proposed AE architecture for feature reduction in Fig. 4 and provides detailed parameter information for the proposed AE in Table IV.Source: Figure 4 · p. 7, Table IV · p. 7

1) Standard ANFIS:

  • Standard ANFIS uses a first-order Sugeno fuzzy model with IF-THEN rules and adaptive membership-function parameters.Normalized rule firing strengths weight first-order polynomial outputs, which the final fixed node sums.Source: 1) Standard ANFIS · pp. 7–8

2) Optimal ANFIS Classifiers:

The ANFIS variants replace backpropagation training with optimization algorithms that tune membership-function parameters to minimize a cost function. Breeding Swarm (BS) combines genetic-algorithm and particle-swarm steps.

2) Optimal ANFIS Classifiers:

  • PSO and BS optimize ANFIS membership-function parameters to minimize a cost function, while GOA, PSO, and BS replace backpropagation for network training.The cost function uses ANFIS error, inputs, parameters, and outputs; the passage identifies N as the number of inputs and n as the number of instances.Source: 2) Optimal ANFIS Classifiers · p. 8
  • BS combines genetic-algorithm selection, crossover, and mutation with particle-swarm position and velocity updates, supporting global and local search, respectively.The procedure generates a population, evaluates fitness, selects P best particles by roulette wheel, and runs GA and PSO steps in parallel.Source: 2) Optimal ANFIS Classifiers · pp. 8–9
  • The flowchart depicts the implementation steps of the BS algorithm.The figure caption identifies the flowchart's scope but does not specify additional algorithm details.Source: Figure 6 · p. 9

III. STATISTICAL PARAMETERS FOR CLASSIFICATION PERFORMANCE

The experiments compare fuzzy and non-fuzzy classifiers on Bonn and Freiburg EEG datasets, reporting average results from ten runs. ANFIS-BS achieves the highest reported classifier performance, while the authors note computational and configuration limitations.

III. STATISTICAL PARAMETERS FOR CLASSIFICATION PERFORMANCE

  • Classification performance was evaluated with 10-fold cross-validation using specificity, sensitivity, accuracy, and F1-score.The metrics are derived from true-positive, false-negative, true-negative, and false-positive counts.Source: §III · p. 9

IV. EXPERIMENT RESULTS

  • The experiments use 5-second Bonn windows and 4-second Freiburg windows, with TQWT parameters Q = 1, r = 3, and J = 8 for both datasets.The same preprocessing steps are applied across datasets; TQWT decomposes the signals into frequency sub-bands.Source: §IV · p. 9
  • Fuzzy-entropy execution times vary: FuEn is fastest and FuDistEn slowest, while AE reduction changes 135 extracted features to 32.The reported feature count comes from 15 fuzzy entropies across nine TQWT sub-bands; the runtime comparison includes raw EEG and TQWT sub-bands.Source: §IV · p. 10, Figure 7 · p. 10
  • ANFIS-BS outperformed the other classification methods in average results for the Bonn and Freiburg datasets.Each method was run ten times under identical conditions; the Bonn classifier results are summarized in Figure 9, and Freiburg results in Table VII.Source: §IV · p. 10, Table VII · p. 11

V. LIMITATIONS OF THE STUDY

  • The study identifies feature-extraction computational cost and trial-and-error fuzzy-entropy parameter selection as limitations, and notes that its TSK implementation uses Gaussian membership functions only.The authors also state that ANFIS use and tuning PSO and BS parameters require optimization and fuzzy-systems knowledge.Source: §V · p. 10

VI. DISCUSSION, CONCLUSION AND FUTURE DIRECTIONS

The paper evaluates a fuzzy-entropy and deep-learning EEG detection pipeline on Bonn and Freiburg datasets, using TQWT sub-bands, a six-layer autoencoder, and several classifiers. ANFIS-BS reports the highest accuracy among the tested fuzzy and non-fuzzy classifiers, and the authors propose broader clinical and seizure-prediction evaluations as future work.

VI. DISCUSSION, CONCLUSION AND FUTURE DIRECTIONS

  • The proposed pipeline windows EEG signals and decomposes them into TQWT sub-bands before extracting type-1 fuzzy entropies.Bonn uses 5-second windows and Freiburg 4-second windows; both datasets are decomposed into nine sub-bands.Source: §VI · pp. 11–12
  • The authors present fuzzy-entropy feature extraction and its computational-time complexity analysis as novelties, alongside a six-layer AE architecture for dimensionality reduction.The paper also presents runtime comparisons for fuzzy entropies as relevant to practical applications and hardware implementation.Source: §VI · p. 12
  • ANFIS-BS outperformed ANFIS-PSO and ANFIS-GOA, which the authors identify as first applications of those optimized ANFIS methods to epileptic-seizure detection.The authors attribute the performance advantage to the BS optimizer's combination of GA and PSO.Source: §VI · p. 12
  • ANFIS-BS achieved 99.74% accuracy in Bonn two-class classification, 99.46% in Bonn three-class classification, and 99.28% accuracy on Freiburg.The authors report the highest accuracy among the tested fuzzy and non-fuzzy classifiers, evaluated using 10-fold cross-validation.Source: §VI · p. 12, Figure 9 · p. 12
  • The paper compares its approach with prior AI-based studies on Bonn and Freiburg and reports higher accuracy than most studies.The comparison is presented in Tables VIII and IX; the supplied text does not give their individual comparison values.Source: §VI · pp. 12–13, Table VIII · p. 13, Table IX · p. 14
  • Future work proposed includes evaluation on more complex or clinical datasets, seizure prediction, real-time tools, and additional fuzzy and deep-learning combinations.The authors also suggest smartphone or robotic applications and exploring type-2 fuzzy methods.Source: §VI · pp. 13–14
Loading 2109.04364v2…