Elsevier

Knowledge-Based Systems

Volume 205, 12 October 2020, 106152
Knowledge-Based Systems

A unified multi-level spectral–temporal feature learning framework for patient-specific seizure onset detection in EEG signals

https://doi.org/10.1016/j.knosys.2020.106152Get rights and content

Highlights

  • A novel multi-level spectral–temporal feature learning framework constructed for EEG seizure onset detection.

  • Auxiliary supplementary spectral–temporal information attained.

  • A public dataset with multiple epileptic patients employed to test robustness.

  • High classification performance obtained.

Abstract

Epileptic seizure onset detection in electroencephalography (EEG) signals is a challenging task due to the severe variation of seizures. Recently, automatic seizure onset detection frameworks fail to fully consider both nonstationary and stochastic characteristics of EEGs in nature, which may lead to information default and further produce suboptimal recognition performance consequently. In this work, we propose a patient-specific seizure onset detection method based on fully exploration of auxiliary supplementary spectral–temporal information in EEG signals. Specifically, prior to feature extraction procedure, EEG signals are firstly decomposed into 5 groups of coefficients at different levels based on the clinical interest. Representative feature in temporal-domain, which is a translation of the nonlinear property of EEG signals, is then extracted by a combination of principal component analysis and common spatial pattern (PCA-CSP) and multivariate multiscale sample entropy (MMSE) in parallel and dimensionally reduced by a tree-based feature selection algorithm. Supplementary information in spectral-domain is further explored by the proposed unified maximum mean discrepancy autoencoder (uMMD-AE). Finally, an optimal combination of features above is identified and fed into a series of support vector machine classifiers with a decision fusion module for the intelligent recognition of epileptic EEGs. The proposed method achieves an average sensitivity, latency and false detection rate of 97.2%, 1.10s and 0.64/h respectively on Children Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) Scalp EEG Database. Competitive experimental results demonstrate the efficacy of the proposed unified multi-level spectral–temporal feature learning framework in epileptic EEG recognition, validating its effectiveness in the automatic patient-specific seizure onset detection.

Introduction

Epilepsy is one of the most common neurological diseases and characterized by excessive and sudden electrical discharge [1]. Generally, manual seizure detection requires long-time scanning of EEG signals, which is commonly time-consuming and error prone [2]. As a result, it is imperative to develop an intelligent and reliable epileptic EEG recognition system for seizure onset detection [3]. However, it is a challenging task due to the stochastic and nonstationary properties of EEGs in nature [4].

Recently, machine learning based approaches have widely gained attention after their immense success in image classification [5], [6], [7], [8], [9]. For intelligent recognition of epileptic EEG signals, several automated machine learning methods have been proposed [3], [4], [10], [11]. Several of these studies firstly extracted discriminative features in temporal-domain, owing to the stochastic characteristic of EEG signals [10], [12]. Particularly, nonlinear methods including the Lyapunov exponent [13], the intrinsic mode functions (IMF) [14] and entropy-based approaches [15], [16] have been adopted to quantify the regularity and unpredictability of fluctuations in epileptic EEGs. For example, Riaz et al. [14] employed the empirical mode decomposition (EMD) to decompose EEG epochs into a set of IMFs, and then temporal-domain features including up to third order temporal moments were extracted for classifying the epileptic EEG signals. Lin et al. [17] adopted features including the approximate entropy for classifying ictal and inter-ictal epochs. However, extracted features in temporal-domain cannot solely provide an effective way for recognizing epileptic EEGs, due to its neglecting the highly nonstationary property of EEG signals. This may undermine the recognition performance to some extent.

Apart from temporal-domain feature extraction, analyses in spectral-domain have also received considerable attention for the automatic seizure detection, considering its capability of translating the nonstationary characteristic of EEG signals [18], [19], [20]. For instance, Faust et al. [13] deployed EEG denoising and the feature extraction by using a discrete wavelet transform (DWT) method. A tunable factor wavelet transform scheme was adopted for the automatic recognition and detection of epileptic EEGs [21]. However, spectral-domain features extracted based on a fixed frequency range neglect the heterogeneity among epileptic patients, which may undermine the performance of above studies in patient-specific seizure detection and diagnosis scenario [22]. The scalogram based convolutional neural network (CNN) could tackle this problem to some extent due to its data-driven training mode [23], [24], [25]. For example, Yuan et al. [26] integrated a CNN with convolutional autoencoders (cAE) for multi-view spectrum representation learning using the short time Fourier transform (STFT) method. However, above studies achieve a limited success with failure to adopt spectral–temporal features but only restricted in single-domain consideration. Therefore, it is worth proposing a multi-domain feature learning framework for the better recognition and detection of epileptic EEGs.

Although multi-domain feature extraction schemes fully consider both stochastic and nonstationary characteristics of EEGs simultaneously [27], [28], it is far from satisfactory to construct parallel-stream architectures for each domain respectively [27], [29] or extract features in a domain-agnostic manner [30], [31]. Recently, maximum mean discrepancy [32], which is a prevalent distribution discrepancy measurement, achieves a good performance in the transfer learning and unsupervised learning. We further explore its ability in the multi-domain feature heterogeneity resolution and construct a unified multi-domain feature learning framework to overcome the variation among multi-domain information consequently.

To address aforementioned issues, in this study, we propose an intelligent recognition scheme of epileptic EEGs to automatically detect seizure onsets by using a unified multi-level spectral–temporal feature learning framework. Specifically, prior to the feature extraction, EEG signals are firstly decomposed into multi-level coefficients based on the clinical interest. Secondly, multi-domain feature extraction branches into 5 parallel subnets, corresponding to 5 groups of EEG wavelet coefficients respectively. Discriminative features in temporal-domain are firstly extracted by using a combination of principal component analysis and common spatial pattern (PCA-CSP) and multivariate multiscale sample entropy (MMSE) respectively. Then a tree-based feature selection algorithm is applied to generate an optimal subset of representative temporal-domain features. Furthermore, a unified maximum mean discrepancy autoencoder (uMMD-AE) is adopted for the supplementary spectral-domain feature extraction, which takes spectral–temporal complementarity into account simultaneously. Finally, a series of support vector machine (SVM) classifiers with a decision fusion module are employed for epileptic EEG recognition. Experimental results on CHB-MIT dataset demonstrate that our proposed method can well capture spectral–temporal features embedded in raw EEGs and detect seizure onsets more efficiently against the state-of-the-art methods.

One significant contribution of the proposed scheme is that a unified framework of multi-level spectral–temporal feature learning is applied to integrate statistical and morphological features, non-linear analysis and spectral based deep learning algorithm together, which considers the complementary relationship between multi-domain features as well. Another main contribution is that a series of SVM classifiers with a decision fusion module are adopted to produce more reliable classification results, which mitigates the overfitting problem caused by the scarcity of seizure events. As a result, a consistent improvement with different patients is achieved by our proposed scheme for recognizing and detecting epileptic EEG signals effectively.

Section snippets

Database

In this paper, a public epileptic EEG dataset is employed to evaluate the effectiveness of the proposed seizure onset detection approach. Specifically, CHB-MIT dataset [33], collected at Children’s Hospital Boston, consists of continuous multi-channel EEG recordings at a sampling rate of 256 Hz from 23 pediatric patients. Recordings are grouped into 24 cases, each of which has EEG signals from a single subject. All the EEG recordings were collected based on the protocol of international 10–20

Methodology

A detailed flowchart of the proposed seizure onset detection framework is shown in Fig. 1. Specifically, EEG segments are firstly preprocessed with baseline removal, detrend and lowpass filter (0–64 Hz) operations for the noise and artifact removal. Five groups of wavelet coefficients at different levels, indicating δ (0–4 Hz), θ (4–8 Hz), α (8–16 Hz), β (16–32 Hz) and γ (32–64 Hz) rhythms respectively, are further generated using DWT method. Next, multi-domain feature extraction is employed

Overall performance

For the comparison and clarity, a segment of EEG recording of a specific patient is shown in Fig. 3 with only the 5 selected channels given. From Fig. 3, all of selected channels exhibit significant fluctuations within the ictal stage compared with the inter-ictal duration. Additionally, wavelet decompositions of a 4s EEG epoch of the inter-ictal class and ictal class are shown in Fig. 4(a) and (b) respectively, where we plot amplitude of wavelet coefficients on the vertical Y-axis against time

Discussions

As is claimed in Section 3, the preprocessing procedure adopted in this work does not guarantee a real-time seizure onset detection but theoretically can serve as an off-line seizure annotation system. With this aside, we test the real-time performance of the proposed method, which is one of the most important factor of a seizure onset detection framework. We implement the proposed method using MATLAB version R2018b and the computations are performed on a standard desktop computer with a

Conclusions

In this work, a novel intelligent recognition framework of epileptic EEGs is proposed for automatic seizure onset detection by means of a unified multi-level multi-domain feature learning framework. Specifically, a multi-level EEG framework is constructed with the DWT algorithm. Statistical and morphological features and nonlinear features in temporal-domain are firstly extracted in parallel. Complementary spectral-domain features are then learned in an unified manner, aided by the proposed

CRediT authorship contribution statement

Fang-Gui Tang: Conceptualization, Methodology, Software. Yu Liu: Conceptualization, Methodology, Software. Yang Li: Data curation, Writing - original draft. Zi-Wen Peng: Resources, Supervision, Project administration.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (46)

  • WangL. et al.

    Automatic epileptic seizure detection in EEG signals using multi-domain feature extraction and nonlinear analysis

    Entropy

    (2017)
  • AcharyaU.R. et al.

    An integrated index for detection of sudden cardiac death using discrete wavelet transform and nonlinear features

    Knowl.-Based Syst.

    (2015)
  • KiranyazS. et al.

    Automated patient-specific classification of long-term electroencephalography

    J. Biomed. Inform.

    (2014)
  • FuK. et al.

    Classification of seizure based on the time-frequency image of EEG signals using HHT and SVM

    Biomed. Signal Process. Control

    (2014)
  • OroscoL. et al.

    Patient non-specific algorithm for seizures detection in scalp EEG

    Comput. Biol. Med.

    (2016)
  • BhattacharyyaA. et al.

    A multivariate approach for patient specific EEG seizure detection using empirical wavelet transform

    IEEE Trans. Biomed. Eng.

    (2017)
  • ZabihiM. et al.

    Analysis of high-dimensional phase space via Poincaré section for patient-specific seizure detection

    IEEE Trans. Neural Syst. Rehabil. Eng.

    (2015)
  • LiY. et al.

    Epileptic seizure detection based on time-frequency images of EEG signals using Gaussian mixture model and gray level co-occurrence matrix features

    Int. J. Neural Syst.

    (2018)
  • DengZ. et al.

    Transductive joint-knowledge-transfer TSK FS for recognition of epileptic EEG signals

    IEEE Trans. Neural Syst. Rehabil. Eng.

    (2018)
  • ChenC.L.P. et al.

    I-ching divination evolutionary algorithm and its convergence analysis

    IEEE Trans. Cybern.

    (2017)
  • ZhangT. et al.

    Hierarchical lifelong learning by sharing representations and integrating hypothesis

    IEEE Trans. Syst. Man Cybern. Syst.

    (2019)
  • ChenH. et al.

    An enhanced bacterial foraging optimization and its application for training kernel extreme learning machine

    Appl. Soft. Comput.

    (2019)
  • ChenH. et al.

    Efficient multi-population outpost fruit fly-driven optimizers: Framework and advances in support vector machines

    Expert Syst. Appl.

    (2019)
  • Cited by (31)

    • 3D residual-attention-deep-network-based childhood epilepsy syndrome classification

      2022, Knowledge-Based Systems
      Citation Excerpt :

      Compared with conventional machine-learning methods, deep networks, which benefit from the promising learning capability in complex data, have been extensively studied in epilepsy analysis [11–18]. However, seizure onset detection has been the main focus in current research [19,20], and feature extraction has generally been conducted by using two-dimensional (2D) time–frequency EEG spectrograms. Recently, 3D convolutional deep network (C3D) [21], residual ConvNet (R3D) [22], and Expand 3D (X3D) [23] have become popular for EEG-based seizure detection [24–27].

    • Hierarchical Harris hawks optimization for epileptic seizure classification

      2022, Computers in Biology and Medicine
      Citation Excerpt :

      The reason for this result was that the data amount was too small. Refs. [1,6,85] only used five channels, and their classification accuracies were 99.41%, 95.96%, and 97.8%, respectively. The latter two were used to extract frequency-domain features and time-domain features.

    • Complex network modeling of EEG band coupling in dyslexia: An exploratory analysis of auditory processing and diagnosis

      2022, Knowledge-Based Systems
      Citation Excerpt :

      They extract features from the EEG signals and classify them. In [20] a SVM is trained with spectral–temporal features for EEG seizure onset detection, and Darjani and Omranpour [21] extract features from the phase space of the EEG signals for epileptic seizure classification. EEG has also been successfully applied to DD.

    • Seizure detection from multi-channel EEG using entropy-based dynamic graph embedding

      2021, Artificial Intelligence in Medicine
      Citation Excerpt :

      In order to compare the proposed method to the previous, we utilized four seizure detection methods applied on the same database as baselines, which are shown as follows. Tang et al. [23] presented a multi-level spectral-temporal feature learning framework for seizure detection. It decomposed the multi-channel EEG into five groups, extracted the groups' features, and used a support vector machine as a classifier to detect the seizure.

    View all citing articles on Scopus

    This document is by the National Natural Science Foundation of China [U1809209, 61671042, 61403016, 31871113], Beijing Natural Science Foundation, China [L182015, 4172037], the Zhejiang Provincial Natural Science Foundation of China [LSZ19F020001], the Major Project of Wenzhou Natural Science Foundation, China [ZY2019020], and the National Natural Science Foundation of China [31871113, 31920103009].

    View full text