ABSTRACT

Electroencephalography (EEG) plays a vital role in detecting how brain responses to different stimulus. In this paper, we propose a novel Shallow-Deep Attention-based Network (SDANet) to classify the correct auditory stimulus evoking the EEG signal. It adopts the Attention-based Correlation Module (ACM) to discover the connection between auditory speech and EEG from global aspect, and the Shallow-Deep Similarity Classification Module (SDSCM) to decide the classification result via the embeddings learned from the shallow and deep layers. Moreover, various training strategies and data augmentation are used to boost the model robustness. Experiments are conducted on the dataset provided by Auditory EEG challenge (ICASSP Signal Processing Grand Challenge 2023). Results show that the proposed model has a significant gain over the baseline on the match-mismatch track.

INTRODUCTION

ICASSP 2023 Auditory EEG Challenge is designed to explore the relationship between auditory stimulus and evoked EEG signal. In this paper, we mainly focus on the first (match-mismatch) task. Traditional methods [1] adopt the linear model to fit the feature transform from the stimulus to the EEG signal. Recently, deep learning based methods have been proposed to improve the performance of relating speech with EEG signal[2, 3], where the feature transform modules are replaced by long context model, such as long short-term memory (LSTM), stacked dilated Convolutional blocks, etc.

CONCLUSION

The objective of the study is to learn the relationship between auditory stimulus and EEG. We propose a SDANet model with ACM and SDSCM to increase the crossing linkages rather than treating them separately. Results show that our model improves the classification accuracy effectively.

Abstract

     Patients with anxious depression have more severe symptoms, more side effects, and higher resistance to treatment than patients with non-anxious depression; therefore, it is crucial to clarify the differences between these two types of patients. In this study, a 5-minute resting EEG was recorded in 15 patients with anxious depression and 9 patients with non-anxious depression under eyes open and closed conditions. Sixty-eight subcortical regions were extracted using exact low resolution brain electromagnetic tomography (eLORETA). The directed transfer function was then used to construct brain networks. Specific features based on graph theory including the strength of connectivity and betweenness centrality (BC) were calculated from the networks. Finally, significant features were selected using the Mann-Whitney U test, and patients were classified into anxious and non-anxious depressive groups using the Support Vector Machine (SVM). Results showed that features of outward connectivity strength led to the highest accuracy, F-score, and specificity with 91.66%, 87. 5%, and 100% in the eyes-closed state, respectively. Moreover, we found that the strength of connectivity in both directions increased for the anxious depressive group during the eyes-open state. In particular, higher outward connectivity was observed in the right hemisphere for the anxious depressive group. Further findings also revealed that features with the most significant difference were mainly associated with the beta band. In addition, significant increased inward and outward connectivity and decreased nodal centrality were observed in the posterior regions of the default mode network. These preliminary findings might provide new insights into the recognition of anxious depressed patients.

Introduction

     Major depressive disorder (MDD) affects 6% of the global adult population annually and is one of the most prevalent psychiatric disorders [1]. About half of MDD patients also have anxiety, such as social anxiety disorder (SAD), generalized anxiety disorder (GAD), or panic disorder (PD) [2]. When MDD co-occurs with anxiety, the risk of suicide becomes greater, and patients may require long-term treatment [3,4]. Currently, however, there is a lack of accepted treatment for patients with anxious depression. They usually receive the same treatment strategies that are given for depression or anxiety. Consequently, the rate of treatment resistance is higher for them than for patients with pure depression or anxiety [5]. Moreover, the rate of medical utilization increased rapidly for patients with comorbid depression and anxiety [4]. Therefore, it is necessary to distinguish anxious depressed patients from non-anxious depressed patients in order to improve diagnosis and treatment methods.

Conclusion

     To the best of our knowledge, this is the first study that has investigated differences in the directed brain network in anxious and nonanxious depressed patients using an effective connectivity measure and the EEG source connectivity method. Network metrics comprising directed node strength and BC were analyzed by a statistical test and the machine learning approach. Classification results demonstrate outward connectivity strength had the highest performance in separating the two groups of patients, while all features performed better in the eyes-closed state than in the eyes-open state. Our main findings related to the strength of connectivity consist of increased node strength during the eyes-open condition, especially a higher out-strength in the right hemisphere for the anxious depressive group. In addition, we found that beta oscillations reflect the most altered brain network in terms of node strength and BC. Further analysis revealed that connectivity and centrality in most regions of posterior DMN were changed significantly. To conclude, the obtained results could potentially lead to the understanding of the underlying brain network of patients with comorbid depression-anxiety disorder, however, more research still is needed regarding the anxious depressive disorder.

Abstract

Electroencephalography (EEG), which tracks the brain waves that contain the brain’s neural activity, plays an essential role in detecting human motion and treating neurological diseases. In the Artificial Intelligence (AI) era, deep learning algorithms are widely used in human action recognition and classification. Various convolutional neural networks that process this signal are also being born. This paper provides a detailed survey of the application of deep learning to EEG signals and outlines the research process when classifying EEG signals. At the same time, this paper reviews the relevant research on the classification of human action EEG signals in recent years. Human motion signals usually use different deep learning algorithms and convolutional neural network architectures in the EEG signal analysis task. This article will discuss the advantages and challenges of each method in other studies. Finally, the paper discusses future directions for deep learning-based EEG signal classification.

Abstract

Depression is a disorder that if not treated can hamper the quality of life. EEG has shown great promise in detecting depressed individuals from depression control individuals. It overcomes the limitations of traditional questionnaire-based methods. In this study, a machine learning-based method for detecting depression among young adults using EEG data recorded by the wireless headset is proposed. For this reason, EEG data has been recorded using an Emotiv Epoc+ headset. A total of 32 young adults participated and the PHQ9 screening tool was used to identify depressed participants. Features such as skewness, kurtosis, variance, Hjorth parameters, Shannon entropy, and Log energy entropy from 1 to 5 sec data filtered at different band frequencies were applied to KNN and SVM classifiers with different kernels. At AB band (8–30 Hz) frequency, 98.43 ± 0.15% accuracy was achieved by extracting Hjorth parameters, Shannon entropy, and Log energy entropy from 5 sec samples with a 5-fold CV using a KNN classifier. And with the same features and classifier overall accuracy = 98.10 ± 0.11, NPV = 0.977, precision = 0.984, sensitivity = 0.984, specificity = 0.976, and F1 score = 0.984 was achieved after splitting the data to 70/30 ratio for training and testing with 5-fold CV. From the findings, it can be concluded that EEG data from an Emotiv headset can be used to detect depression with the proposed method.

Introduction

Depressive disorder is a highly prevalent mental illness. Sadness, loss of interest or enjoyment, feelings of guilt or low self-worth, interrupted sleep or food, fatigue, and difficulty concentrating are some characteristics of depression. It may affect a person’s capacity to operate in daily life or at work or school. According to the World Health Organization (WHO) back in 2015, almost 4.4% of the world’s population was suffering from depression [1]. Because of the COVID-19 pandemic, many people suffered from depression due to job loss, study hampering, losing close relatives, staying indoors, etc. A study showed 19.3% increase in depression symptoms among people after COVID-19 in the United States [2]. A study has shown the changes in obsession, depression, and quality of life in schizophrenia patients before and after COVID-19 [3]. When depression is severe it can lead to suicide. Every year around 800 thousand people die because of suicide [1]. In 2017, 13.2% of young adults (aged 18–25) in the U.S. suffered from depression which was 5.1% less in the year 2009 [4]. Of the deaths of young people, around 9.1% are due to suicide [5]. In most suicide cases, people had psychiatric disorders where depression is the most common disorder among others [6]. According to a recent study, insecure attachment styles are linked to greater problems such as depression, social anxiety, and suicidal thoughts [7]. So, depression is a major issue that should be diagnosed and treated at an early age to prevent suicide and for the betterment of the quality of life.

Conclusion

In this work, we have recorded EEG data of young adults (19 depressed and 13 Control) evaluated by the PHQ9 screening tool and proposed a machine learning approach to learn about the EEG properties for depression detection.

We conducted multiple experiments with the reported machine learning (SVM and KNN) classifiers with our recorded data. The first experiment we conducted was on segmentation to find the better sample length suitable for ML. From our experiments, we have identified that 5-second segments are suitable for our work. Then, we have identified a suitable frequency range from various experiments that improve performance using features that are related to depression detection. We have found out that Hjorth parameters along with Shannon entropy and long energy entropy provide better results among other reported features and the beta band (12–30 Hz) gives the highest accuracy of 97.21 ± 0.21% with 25 iterations and 5-fold CV using weighted KNN compared to the other sub-bands. By combining the sub-bands, we have also investigated some other frequency ranges. We have found out that by taking the range from alpha to beta 8–30 Hz (AB), we can improve ML performance and achieve 98.43 ± 0.15% accuracy with 25 iterations and a 5-fold CV with fine KNN classifier. Using AB (8–30 Hz), we can see a significant improvement of 1.22% accuracy and slandered deviation. To further investigate the reliability, we divided the dataset 70/30 for training and testing with 5-fold CV and 10 iterations. In this experiment, we have found out that the ML performance is better by choosing the AB (8–30 Hz) band with fine KNN classifier with an accuracy of 98.10 ± 0.11%, precision of 0.984 ± 0.003, NPV of 0.977 ± 0.002, sensitivity of 0.984 ± 0.002, specificity of 0.976 ± 0.005, and F1 score of 0.984 ± 0.001. Then, we analyzed the ML performance in different regions of the brain and concluded that using the whole brain for depression detection will give the highest accuracy. The proposed method can detect depression among young adults with minimum requirements compared to other related works.

Abstract

1. Introduction

2. Materials and Methods

3. Proposed Epilepsy Classification Method

4. Results and Discussion

5. Conclusions

Data Availability

Conflicts of Interest

References

Abstract

     The precise detection of epileptic seizure helps to prevent the serious consequences of seizures. As the electroencephalogram (EEG) reflects the brain activity of patients effectively, it has been widely used in epileptic seizure detection in the past decades. Recently, deep learning-based detection methods which automatically learn features from the EEG signals have attracted much attention. However, with deep learning-based detection methods, different input formats of EEG signals will lead to different detection performances. In this paper, we propose a deep learning-based epileptic seizure detection method with hybrid input formats of EEG signals, i.e., original EEG, Fourier transform of EEG, short-time Fourier transform of EEG, and wavelet transform of EEG. Convolutional neural networks (CNNs) are designed for extracting latent features from these inputs. A feature fusion mechanism is applied to integrate the learned features to generate a more stable syncretic feature for seizure detection. The experimental results show that our proposed hybrid method is effective to improve the seizure detection performance in few-shot scenarios.

Introduction

     Approximately one percent of the world’s population, 65 million people, suffer from epilepsy, more than Parkinson’s disease, Alzheimer’s disease, and multiple sclerosis combined [1]. About two-thirds of people with epilepsy can be treated with medication, and the rest may require surgical intervention. Epilepsy has the characteristics of sudden and recurrent seizures, which may lead to falls, asphyxia, and even death. Therefore, seizure detection is very important for early warning and treatment of epilepsy.

     Epileptic seizure detection is mainly based on electroencephalogram (EEG) [2–4]. Single-channel EEG acquisition equipment improves the practicability of EEG in epileptic detection due to its simplicity in implementation. However, the provided information by signal-channel EEG signal is limited because of the small number of channels. Therefore, it is worth studying to establish a model with high accuracy and high robustness for single-channel EEG epileptic detection.

     The traditional methods are mainly based on feature engineering techniques which extract the corresponding features from EEG signals and then complete the detection based on the extracted features [5–9]. These features include time-domain features [10–12], frequency-domain features [8, 9], and time-frequency-domain features [13–15]. Once the features are extracted, EEG signals can be classified using a variety of classifiers. No matter what classifier is used, the quality of designed features will greatly affect the performance of epilepsy detection. In recent years, with the development of deep learning technology, many works have applied deep learning to perform epilepsy detection [16, 17]. Different from traditional feature engineering, deep learning methods automatically learn features from EEG signals and further complete detection tasks with an end-to-end manner without complicated manual feature design process and can achieve better performance than traditional methods in many scenarios.

Conclusions

     In this paper, we focus on epileptic classification in few-shot scenarios. In order to make the classification accuracy higher and more stable, we propose a deep learning method with hybrid input, i.e., original EEG signal, FFT, STFT, and DWT of EEG signal. In order to alleviate the tendency of overfitting, two means are applied. The first is that we replace the traditional convolution by depthwise separable convolution for reducing the parameters in network and then 2 regularization is applied, whose function is decreasing the complexity of the model. We conduct several experiments to distinguish normal and epileptic EEG, and the results show the proposed method with hybrid input has strong advantages in epileptic classification. It benefits from the complementarity of time-domain properties, frequency-domain properties, and time-frequency-domain properties. Our proposed method provides a new perspective to enrich the input information to make improvements for deep learning-based epileptic diagnosis.

Highlights

Abstract

Keywords

Introduction

Method

Results

Discussion

Conclusions

Acknowledgments

Author contributions

Appendix A. Supplementary material

References

Abstract

Did I make that sound? Differentiating whether sensory events are caused by us or the environment is pivotal for our sense of agency. Adults can predict the sensory effects of their actions, which results in attenuated processing of self-produced events compared with externally generated events. Yet, little is known about whether young infants predict and discriminate self-produced events from externally produced events. Using electroencephalography (EEG), 3-month-olds’ neural response to the same audiovisual stimulus was compared between a Self-produced condition and externally generated conditions with predictable timing (External–Regular) and irregular timing (External–Irregular). We hypothesized that if 3-month-olds predict self-produced events, their event-related potentials should be smallest for the Self-produced condition, strongest for the External–Irregular condition, and in between for the External–Regular condition. Cluster-based permutation tests indicated a more positive deflection (300–470 ms) for irregular stimuli compared with regular stimuli over the vertex. Contrasting the Self-produced and External–Irregular conditions showed a statistical trend within the same time window. Although not fully conclusive, this might suggest the emerging differentiation between self-produced and less predictable external events. However, there was no statistical evidence that infants differentiated self-produced events from temporally predictable external events.

Introduction

The sense of agency is the sense that one’s own actions cause effects in the environment. How the sense of agency develops during infancy has long fascinated developmental scientists, philosophers, and new parents alike. The sense of agency is ubiquitous during adulthood and is fundamental for becoming an intentional agent during early childhood, likely forming the basis for infants to learn from and about other intentional agents (Meltzoff, 2007). One defining element of the sense of agency is the ability to distinguish one’s own actions and their effects from external events in the world (Tsakiris, Schütz-Bosbach, & Gallagher, 2007). Research with adults suggests that predicting the sensory effects of one’s actions plays an important role in discriminating self-produced events from externally generated events (e.g., Blakemore, Wolpert, & Frith, 2000; Hughes, Desantis, & Waszak, 2012). As such, the phenomenon that we cannot tickle ourselves illustrates not only that we perceive selfproduced effects differently from other sensory input but also that we process the sensory effects of our own actions in an attenuated fashion. This phenomenon can be explained by our precise prediction of the tactile input when trying to tickle ourselves that attenuates the tickling sensation (Blakemore et al., 2000).

1- Introduction

2- Materials and methods

3- Results

4- Discussion

5- Conclusion

References

Abstract

Gamma rhythms (~20–70 ​Hz) are abnormal in mental disorders such as autism and schizophrenia in humans, and Alzheimer’s disease (AD) models in rodents. However, the effect of normal aging on these oscillations is unknown, especially for elderly subjects in whom AD is most prevalent. In a first large-scale (236 subjects; 104 females) electroencephalogram (EEG) study on gamma oscillations in elderly subjects (aged 50–88 years), we presented full-screen visual Cartesian gratings that induced two distinct gamma oscillations (slow: 20–34 ​Hz and fast: 36–66 ​Hz). Power decreased with age for gamma, but not alpha (8–12 ​Hz). Reduction was more salient for fast gamma than slow. Center frequency also decreased with age for both gamma rhythms. The results were independent of microsaccades, pupillary reactivity to stimulus, and variations in power spectral density with age. Steady-state visual evoked potentials (SSVEPs) at 32 ​Hz also reduced with age. These results are crucial for developing gamma/SSVEP-based biomarkers of cognitive decline in elderly.

Introduction

Gamma rhythms are narrow-band oscillations often observed in the electrical activity of the brain, with center frequency occupying ~20-70 Hz frequency range. Previous studies have proposed involvement of these rhythms in certain higher cognitive functions like feature binding (Gray et al., 1989), attention (Chalk et al., 2010; Gregoriou et al., 2009) and working memory (Pesaran et al., 2002). Further, some studies have shown that these rhythms may be abnormal in neuropsychiatric disorders such as schizophrenia (Hirano et al., 2015; Tada et al., 2014), autism (An et al., 2018; Uhlhaas and Singer, 2007; Wilson et al., 2007) and Alzheimer’s disease (Mably and Colgin, 2018; AD; Palop and Mucke, 2016). Gamma rhythms can be induced in the occipital areas by presenting appropriate visual stimuli such as bars and gratings, and their magnitude and center frequency critically depend on the properties of the stimulus such as contrast, size, orientation, spatial frequency and drift rate (Jia et al., 2013; Murty et al., 2018; Ray and Maunsell, 2015). Recently, we showed that large (full-screen) gratings induce two distinct narrow-band gamma oscillations in local field potentials (LFP) in macaque area V1 and posterior electrodes in human EEG, which we termed slow (~20-40 Hz) and fast (~40-70 Hz) gamma (Murty et al., 2018).

1- Introduction

2- Methodology

3- Results and discussion

4- Conclusions

References

Abstract

Epilepsy is the fourth most common neurological disorder that manifests itself through unprovoked seizures, detection of which is the very first step of proper diagnosis and treatment of this severe disease. In this paper, an automated seizure detection method has been proposed based on the statistical and spectral features of max normalized intrinsic mode functions or IMFs that were extracted using complete ensemble empirical mode decomposition with adaptive noise method. First, a publicly available dataset of EEG signals was used to generate the IMFs and noise or outliers were discarded. Then IMFs were max normalized which was shown to improve the separability of features. Statistical and spectral features were extracted from the normalized IMFs which offered better separation of seizure and seizure-free data. Finally, Quadratic Discriminant classifier was used for the classification purpose and 10-fold cross validation was performed to validate the trained model. The proposed scheme is numerically efficient and shows a maximum of 100% accuracy which is the highest reported on this data set.

Motivation

The International League Against Epilepsy (ILAE) and the International Bureau for Epilepsy (IBE) have defined the terms epileptic seizure and epilepsy as follows: an epileptic seizure is a transient occurrence of signs and/or symptoms due to abnormal excessive or synchronous neuronal activity in the brain and epilepsy is a disorder of the brain characterized by an enduring predisposition to generate epileptic seizures and by the neurobiologic, cognitive, psychological, and social consequences of this condition [1]. The definition of epilepsy requires the occurrence of at least one epileptic seizure. According to World Health Organization (WHO), approximately 50 million people have epilepsy worldwide. Around 80% of epilepsy patients are either from low or middleincome countries, of which around 75% failto get proper treatment. Epilepsy has severe socioeconomic consequences, including high treatment cost,low employment rate, andlow salary among epileptic patients [2]. Epilepsy remains resistant to drug therapy in about one-third of patients and people with pharmacoresistant epilepsy are about 2–10 times more likely to die compared to the general population [3].

1- Introduction

2- Related work

3- Materials and method

4- Experimental results

5- Discussion

6- Conclusion and future direction

References

Abstract

Epilepsy, a common neurological disorder, is generally detected by electroencephalogram (EEG) signals. Visual inspection and interpretation of EEGs is a slow, time consuming process that is vulnerable to error and subjective variability. Consequently, several efforts to develop automatic epileptic seizure detection and classification methods have been made. The present study proposes a novel computer aided diagnostic technique (CAD) based on the discrete wavelet transform (DWT) and arithmetic coding to differentiate epileptic seizure signals from normal (seizure-free) signals. The proposed CAD technique comprises three steps. The first step decomposes EEG signals into approximations and detail coefficients using DWT while discarding non-significant coefficients in view of threshold criteria; thus, limiting the number of significant wavelet coefficients. The second step converts significant wavelet coefficients to bit streams using arithmetic coding to compute the compression ratio. In the final step, the compression feature set is standardized, whereupon machine-learning classifiers detect seizure activity from seizure-free signals. We employed the widely used benchmark database from Bonn University to compare and validate the technique with results from prior approaches. The proposed method achieved a perfect classification performance (100% accuracy) for the detection of epileptic seizure activity from EEG data, using both linear and non-liner machine-learning classifiers. This CAD technique can thus be considered robust with an extraordinary detection capability that discriminates epileptic seizure activity from seizure-free and normal EEG activity with simple linear classifiers. The method has the potential for efficient application as an adjunct for the clinical diagnosis of epilepsy.

Introduction

Epilepsy is a common brain disorder that affects people of all ages. It is a chronic neurological disorder in which recurrent seizures occurs due to abnormal neuronal activities within the human brain and affects the sensorium, mood and/or movement of the human body [1]. World Health Organization statistics indicate that approximately 50 million people currently live with epilepsy worldwide and an estimated 2.4 million people are diagnosed with epilepsy each year [2]. The incidence of the malady is higher in developing countries;i.e., between 7 and 14 per 1000 people. Treatment mostly comprises antiepileptic drugs and/or surgery [3]. The electroencephalogram (EEG) commonly detects seizure activity as it reflects electrophysiological conditions of the brain at a given time [4] and is widely used for diagnostic due to its low cost. EEG signals, enhanced with physiological and pathological data, are employed to evaluate and assess the treatment and progress of epileptic patients. Typically, clinicians evaluate EEG signals for three types of activity: (i) normal EEG activity that records healthy subjects with eyes open or closed; (ii) inter-ictal/seizure-free EEG activity that may contain small spikes and/or subclinical seizures that occur between two clinical episodes in epileptic patients; (iii) and ictal EEG activity containing sudden spikes.

1- Introduction

2- Feature extraction

3- Epileptic seizure detection

4- Methods for feature evaluation

5- Experimental results

6- Conclusions

References

Abstract

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. When compared with a baseline method of classifying all epochs as normal, an improvement of 4.77–13.51% in the Bayesian error was obtained.

Introduction

An epileptic seizure, as defined by the International League Against Epilepsy [1], is a temporary event of symptoms due to synchronization of abnormally excessive activities of neurons in the brain. It has been estimated that approximately 65 million people around the world are affected by epilepsy [2]. Nevertheless, it is still a time-consuming process for neurologists to review continuous electroencephalograms (EEGs) to monitor epileptic patients. Therefore, several researchers have developed differenttechniques that help neurologists to identify an epilepsy occurrence [3–5]. The whole process of automated epileptic seizure analysis primarily consists of data acquisition, signal pre-processing, feature extraction, feature or channel selection, and classification. This paper focuses on a selection of features commonly used in the literature, including statistical parameters (mean, variance, skewness, and kurtosis), amplitude-related parameters (energy, nonlinear energy, line length, maximum and minimum values) and entropyrelated measures. These features can be categorized based on their interpretation or the domain from which the features are calculated. While some studies have considered a particular group of features applicable to their proposed classification method [6–8], others have applied various groups of features extracted from the time, frequency, and time-frequency domains. For example, 55 features were used with a support vector machine (SVM) and postprocessing for neonatal seizure detection, which provided a good detectionrate of 89.2% withone false detectionperhour [9]. Feature redundancy and relevance analysis were applied to 132 features to reduce the vector dimension [10]. Fast correlation based-filter proposed in [11], correlation-based selection (CFS) [12], and ReliefF established in [13], were utilized to select non-redundant features and the filtered features were classified by an artificial neural network (ANN). It turned out that 30-optimal selected features by ReliefF achieved the best result with 91% sensitivity and 95% specificity. These studies showed thatthere is the need to review feature selection as relevant features are directly related to the seizure classification performance.

1- Introduction

2- Methods

3- Results

4- Discussion

5- Conclusion

References

Abstract

Introduction: Recent studies showed that neurodevelopment in preterm infants can be predicted by using amplitude-integrated electroencephalography (aEEG)-derived parameters. In our previous study we demonstrated that aEEG could be useful in predicting neurodevelopmental outcome in very preterm infants at the corrected age of 2 years. Aim: The aim of this study was to further evaluate aEEG for predicting neurodevelopmental outcome at the at the corrected age of 2 years in preterm infants. Methods: Between July 2010 and June 2016 440 very preterm infants were eligible for the study at Innsbruck Medical University Hospital. The aEEG was evaluated for the Burdjalov score in 306 preterm infants (mean gestational age 29.5 weeks; range: 24.1–31.9 weeks). At the corrected age of 2 years outcome was assessed by the Bayley Scales of Infant and Toddler Development. Results: The cohort was divided into three subgroups: 248 infants with normal outcome, 40 infants with delayed outcome and 18 infants with abnormal outcome. Burdjalov scores were lower in infants with delayed outcome than in infants with normal outcome and even lower in infants with abnormal outcome. Post-hoc analysis showed significant differences between normal and delayed psychomotor outcome at 18–24 h (5 (3;6) versus 3 (3;5), p = .024), 30–36 h (6 (4;8) versus 4 (4;6), p = .033), 42–48 h (7 (5;8.5) versus 4 (4;7), p = .003), 54–60 h (7 (6;9) versus 5 (4;7), p = .003), 66–72 h (8 (6;9) versus 6.5 (4.25;7.75), p = .027) and week one (8 (7;10) versus 6.5 (5;8), p = .021). Additionally, when comparing normal to abnormal outcome, a significant difference was found at week four (12 (9;12) versus 8 (7;10), p = .024). The Burdjalov score was only predictive for a delayed psychomotor outcome, presenting the highest area under the curve (0.690) at week two of life. Conclusion: We observed differences in aEEG signals and neurodevelopmental outcome at the corrected age of 2 years, especially for psychomotor outcome. The predictive value of the Burdjalov score regarding neurodevelopmental outcome at the corrected age of 2 years in preterm infants was low.

Introduction

About 15 million babies are born before 37 weeks gestational age (preterm) worldwide every year. Mortality in preterm infants has been significantly reduced due to progress in obstetrical and neonatal care. However, morbidity is still of great concern, because preterm infants frequently display neurological sequelae including disturbances in cerebral grey and white matter brain development [1]. In long-term preterm infants suffer from psychomotor and cognitive difficulties, which could be improved by identifying infants at high risk and offering special care and extra resources, such as physical therapy, speech and occupational therapy, and different sorts of education services. The evaluation of brain function complements neurological diagnostic methods used to provide important information in order to identify infants at risk and to offer best care to the preterm infant, to advice and to support the parents and caregivers [2]. Associations of amplitudeintegrated electroencephalography (aEEG)-derived parameters with short-term adverse outcome, such as intracerebral haemorrhage or death, have been reported in preterm infants [3].

1- Introduction

2- Proposed approach

3- Techniques

4- Results

5- Conclusion and future scope

References

Abstract

Electroencephalography (EEG) is a non-invasive method for the analysis of neurological disorders. Epilepsy is one of the most widespread neurological disorders and often characterized by repeated seizures. This paper intends to conduct an iterative filtering based decomposition of EEG signals to improve upon the accuracy of seizure detection. The proposed approach is evaluated using All India Institute of Medical Science (AIIMS) Patna EEG database and online CHB-MIT surface EEG database. The iterative filtering decomposition technique is applied to extract sub-components from the EEG signal. The feature set obtained from each segmented intrinsic mode function consists of 2-D power spectral density and time-domain features dynamic mode decomposition power, variance, and Katz fractal dimension. The Hidden Markov Model (HMM) based probabilistic model has been designed using the above-stated features representing the seizure and non-seizure EEG events. The EEG signal is classified based on the maximum score obtained from the individual feature-based classifiers. The maximum score derived from each HMM classifier gives the final class information. The proposed decomposition of EEG signals achieved 99.60% and 99.74% accuracy in seizure detection for the online CHB-MIT surface EEG database and AIIMS Patna EEG database, respectively.

Introduction

EEG has a wide application in the diagnosis of neurological diseases such as dementia, migraine, and epilepsy [1]. Epilepsy is a non-communicable disease that can affect people of different ages. Nearly 70 million people are affected by epilepsy all over the world, and in India alone, 12 million people are suffering from epilepsy [2]. The unexpected simultaneous activity of groups of neurons is called a seizure. The duration of seizure in continuous EEG signals varies from 1 to 10 seconds. Depending on the seizure condition, the recording duration varies from minutes to hours. At present, the seizure detection is manual, and accuracy depends largely on the doctor’s experience. There is an urgent requirement to develop an efficient algorithm which can detect seizure efficiently from continuous EEG signal. There are different state-of-the-art methods proposed for seizure detection. Discrete wavelet transform (DWT) and empirical mode decomposition (EMD) are widely used techniques for data decomposition. EMD time-domain features and frequency domain features with a linear discriminant analysis classifier are used for seizure classification [3]. Multivariate empirical mode decomposition, instantaneous rate, instantaneous amplitude, and artificial neural network are efficient in seizure detection [4]. EMD intrinsic mode function (IMF) and Higher-order moment features are found to be effective in seizure detection [5]. DWT coefficients are found to be efficient in seizure and non-seizure classification [6]. Energy ratio from wavelet coefficients and ant colony classifier achieved good classification accuracy [7]. Stationary wavelet transform has a good application in seizure detection.

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. When compared with a baseline method of classifying all epochs as normal, an improvement of 4.77–13.51% in the Bayesian error was obtained.

Introduction

An epileptic seizure, as defined by the International League Against Epilepsy [1], is a temporary event of symptoms due to synchronization of abnormally excessive activities of neurons in the brain. It has been estimated that approximately 65 million people around the world are affected by epilepsy [2]. Nevertheless, it is still a time-consuming process for neurologists to review continuous electroencephalograms (EEGs) to monitor epileptic patients. Therefore, several researchers have developed differenttechniques that help neurologists to identify an epilepsy occurrence [3–5]. The whole process of automated epileptic seizure analysis primarily consists of data acquisition, signal pre-processing, feature extraction, feature or channel selection, and classification. This paper focuses on a selection of features commonly used in the literature, including statistical parameters (mean, variance, skewness, and kurtosis), amplitude-related parameters (energy, nonlinear energy, line length, maximum and minimum values) and entropyrelated measures. These features can be categorized based on their interpretation or the domain from which the features are calculated. While some studies have considered a particular group of features applicable to their proposed classification method [6–8], others have applied various groups of features extracted from the time, frequency, and time-frequency domains.