Epileptic seizure classification using level-crossing EEG sampling and ensemble of sub-problems classifier

Highlights

• Effectively combines LCADCs, MASA, and adaptive-rate FIR filter for signal denoising.

• Reduces data size using adaptive-rate techniques, features extraction, and selection.

• Proposes ensemble of sub-problems classification to separate closely related classes.

Abstract

Epilepsy is a disorder of the brain characterized by seizures and requires constant monitoring particularly in serious patients. Electroencephalogram (EEG) signals are frequently used in epilepsy diagnosis and monitoring. A new paradigm of battery packed wearable gadgets has recently gained popularity which constantly monitor a patient’s signals. These gadgets acquire the data and transmit it to the cloud for further processing. Power consumption due to data transmission is a major issue in these devices. Moreover, in a constant monitoring environment, the number of classes to be identified are usually higher and overlapping. Existing techniques either require the entire data to be transmitted, such as in deep learning, or suffer from reduced accuracy. In this context, we propose a new framework for EEG based epilepsy detection which requires a low data transmission while maintaining high accuracy for multiclass classification. At the device-end, we use a preprocessing mechanism that uses adaptive rate sampling, modified activity selection, filtering, and wavelet decomposition to extract only a handful of highly discriminatory features to be transmitted instead of the entire EEG waveform. For multiclass classification, we propose a novel ensemble of sub-problems-based classification paradigm to achieve high accuracy using the reduced data. Our proposed solution shows many-fold increase in computational gains and an accuracy of 100% and 99.38% on the 2-class problem when tested on the popular University of Bonn and CHB-MIT datasets, respectively. An accuracy of 99.6% on 3-class, 96% on 4-class, and 92% on 5-class problems is obtained for the University of Bonn dataset.

Introduction

The human brain controls the coordination of muscles and nerves. Epilepsy, also known as seizure disorder, is a known disorder of the brain characterized predominantly by recurrent and unpredictable interruptions of the normal brain function. These are known as epileptic seizures. While epilepsy is mostly clinically diagnosed, scalp electroencephalogram (EEG) is a widely accepted test for epilepsy detection. EEG is a non-invasive technique that provides high temporal resolution. The signals generated using EEG technique are known as EEG signals or EEG waves. In critical situations, a continuous monitoring of epileptic patients is required, and the EEG signals should be analyzed by neurophysiologists in an uninterrupted manner to observe the functionality of the patient’s cerebral system. Moreover, analysis of multichannel EEG signals is essential for effective identification of epileptic symptoms. This necessitates the recording and analysis of a huge volume of data (Schröder and Ombao, 2019). This is a tiresome process and mistakes in diagnosis can occur due to human fatigue, lack of concentration, physiological artifacts, and noise. Therefore, an automated EEG processing and classification process is needed.

A variety of automated EEG processing techniques have been proposed (Alickovic et al., 2018, Hassan et al., 2020, Qaisar and Subasi, 2020, Zhao et al., 2019), among many others. In these studies, the focus is on identifying a suitable combination of preprocessing, feature extraction, and machine learning based classification algorithms for automated epileptic seizure detection. In (Qaisar and Subasi, 2020), for instance, the noise from EEG signals is removed by using the adaptive-rate Finite Impulse Response (FIR) filtering. Features are extracted by using the Autoregressive Burg (ARB). Classification is done on three classes using Rotation Forest (RoF) algorithm. The authors in (Alickovic et al., 2018) use the Multi Scale Principal Component Analysis (MSPCA), and conditioned signal features are extracted using Wavelet Packet Decomposition (WPD). Classification is done on two and three classes using Support Vector Machines (SVM). A different approach is used in (Zhao et al., 2019) where the noise is first diminished using the Butterworth filter and features mined from subbands using the instantaneous energy extraction approach. The subbands decomposition is attained with Stationary Wavelet Transform (SWT). Two, three and five classes of EEG signals are categorized using the Back-Propagation Neural Network (BPNN) algorithm. Another study (Nishad and Pachori, 2020) reduces the noise by subbands decomposition using a Tunable-Q wavelet transform (TQWT) filters bank. Features are extracted from the subbands by using the Cross-Information Potential (CIP) algorithm and Random Forest (RF) algorithm is used for binary classification. In (Riaz et al., 2015), preconditioning is attained using Empirical Mode Decomposition (EMD). Spectral features are extracted from the conditioned signals and classification is done for two and three classes using SVM.

Discrete wavelet decomposition and deep learning are also popular methods employed in the literature. In (Y. Wang et al., 2017), for instance, noise removal is achieved by using the Discrete Wavelet Transform (DWT) followed by statistical features extraction from each subband. They identify three classes using the Extreme Learning Machine (ELM) algorithm. In (Hassan et al., 2020), noise removal is achieved by using the complete ensemble EMD. Features are mined by using the inverse Gaussian approach. The process is used for identifying both two classes and three classes using the adaptive Boosting classifier. In (Swami et al., 2016), the noise removal with subbands decomposition is carried out by using the Dual-Tree Complex Wavelet Transformation (DTCWT). Statistical features are mined from the subbands, and classification is performed using the General Regression Neural Network (GRNN) algorithm for two classes. In (Li et al., 2019), the Multi Scale Radial Basis Function (MRBF) networks are used for enhancing the EEG signals. Gray level co-occurrence matrix (GLCM) and the Fisher Vector (FV) approaches are used for features extraction. The t-test statistical tool is used for dimension reduction and binary classification is performed using SVM. In (Bhattacharyya et al., 2017), noise reduction and subbands decomposition is achieved by using the TQWT. Features of subbands are extracted by using the quality factor-based entropy. SVM is used for classification. In (Qaisar and Hussain, 2021), EEG signals are conditioned by using the adaptive-rate FIR filtering. Conditioned signals are decomposed in subbands by using the adaptive-rate DWT. Information Gain (IG) based statistical subbands features are extracted. A Bagging algorithm is used for classification of two and three classes. Other algorithms include, for instance, using local binary patterns with wavelets (Khan et al., 2020), EEG detection using CNN (Wei et al., 2019), epileptic seizure identification using Fourier analysis (Mehla et al., 2021), seizure classification using higher order statistics and deep neural networks (Sharma et al., 2020), and stacking ensemble based deep neural network (Akyol, 2020). We refer interested readers to a detailed survey of algorithms and features used for EEG detection to (Siddiqui et al., 2020) and (Shoeibi et al., 2021).

The above-mentioned approaches use sophisticated signal processing and machine learning techniques to classify EEG data. The deep learning models, on the other hand, extract the features as part of the classification. While these approaches have resulted in high classification accuracy, they require the complete data signals and have been shown to perform well on two or three classes that are well separated. With recent advances in automated EEG processing, a new avenue has emerged for mobile healthcare solutions. This facilitates the treatment of chronic disorders by allowing an uninterrupted supervision of patients (Qaisar and Hussain, 2021, Qaisar and Subasi, 2020). These solutions are based on wireless EEG implants connected to health sever or cloud via wireless networks (Tohidi et al., 2019, Zeng et al., 2016). Such EEG wearables are self-powered by batteries and should be light weight and tiny (Bayrakdar, 2019). The main factors in these wearables are the bandwidth requirement and power consumption associated with the data transmission. Due to these reasons, algorithms based on deep neural networks which require transmitting the entire signal may not be suitable.

Another aspect for consideration is the increase in the number of classes in the data. For instance, instead of a simple ictal (seizure) and normal class, we wish to also consider the interictal – the time between the seizures, at both the epileptic and non-epileptic zones, or the signals with the eyes open or closed. Thus, these sub-classes of a more generic class are far more difficult to differentiate and require new and better solutions. Similarly, differentiating between an ictal and inter-ictal signal, particularly in a non-epileptic zone, can be a hard problem. In the case of differentiating between only two classes, specific distinguishing features can be selected. However, for multi-class problems that are closely related, more sophisticated algorithms need to be developed that can classify the data with a low number of features.

Therefore, in this paper, we propose a new framework for epileptic seizure classification which can differentiate between multiple classes, and which is also suitable for wearables. The proposed system incorporates ideas from signal processing, feature selection, and machine learning that are optimized to work with fewer data points and differentiate between overlapping classes. To reduce power consumption, we use adaptive sampling of the EEG signals and feature extraction which results in significantly less data transmission. For multiclass classification, we propose a new mechanism which amalgamates the high accuracy of binary classifiers with ensemble learning using a voting mechanism. The proposed methodology results in low data transmission while retaining high classification accuracy compared to existing state-of-the-art algorithms, including for multiclass problems.

The main contributions in this paper are as follows:

• Effectively combining level crossing analog to digital converters (LCADCs), Modified Activity Selection Algorithm (MASA), adaptive-rate FIR filter-based conditioning for signal processing and noise reduction.

• Reducing the data size by using adaptive-rate DWT based subbands decomposition, statistical features extraction, and mutual information (MI) based feature selection. This results a low number of features that helps lower power consumption and bandwidth reduction.

• A novel ensemble of sub-problems classification strategy is proposed for a more precise classification of EEG signals into multiple, closely related classes.

Majority of the existing automated EEG processing and diagnosis systems are fix-rate, such as in (Alickovic et al., 2018, Hassan et al., 2020, Nishad and Pachori, 2020, Zhao et al., 2019). They are based on the principle of uniform Nyquist sampling theorem [18]. The system is designed while considering the worst scenario for a targeted application. This fix-rate EEG processing approach is not efficient. The system performance in terms of compression and processing efficiency can be enhanced by adjusting its parameters in accordance with the signal variations (Qaisar, 2019). In this context, appealing solutions have been proposed by using the level crossing analog to digital converters (LCADCs) (Antony et al., 2018, Qaisar, 2019, Saeed and Hussain, 2020). In (Qaisar and Subasi, 2020) and (Qaisar and Hussain, 2021), the LCADCs are used for realizing effective epileptic seizure detection. The LCADCs acquire the EEG signals at adaptive-rates and render notable compression and computational effectiveness. Therefore, a significant power consumption reduction is achievable by incorporating the compressive and non-uniform sampling approaches (Mesin, 2016, Zeng et al., 2016).

The advantage of the proposed method is that the ensemble of sub-problems classification strategy works by breaking down the multi-class problem into a set of binary classification subproblems. An ensemble of optimized classifiers is trained for each binary classification. During the testing phase, each classifier is used to predict the class using a one-vs-rest strategy and the final class is selected using a voting mechanism. This results in an overall better prediction of classes that may otherwise be confused by a multi-class classifier.

The rest of the paper is structured as follows: Section 2 details the methods and materials used in the proposed work, such as the signal processing and decomposition techniques, feature extraction and selection, and the classification strategy. In Section 3, we describe the experiments and show the results of the proposed technique. Section 4 presents a discussion of the proposed work and compares it to existing strategies. Finally, we conclude the work in Section 5.