Evangelia Pippa

Improving classification of epileptic and non-epileptic EEG events by feature selection

Correctly diagnosing generalized epileptic from non-epileptic episodes, such as psychogenic non epileptic seizures (PNES) and vasovagal or vasodepressor syncope (VVS), despite its importance for the administration of appropriate treatment, life improvement of the patient, and cost reduction for patient and healthcare system, is rarely tackled in the literature. Usually clinicians differentiate between generalized epileptic seizures and PNES based on clinical features and video-EEG. In this work, we investigate the use of machine learning techniques for automatic classification of generalized epileptic and non-epileptic events based only on multi-channel EEG data. For this purpose, we extract the signal patterns in the time domain and in the frequency domain and then combine all features across channels to characterize the spatio-temporal manifestation of seizures. Several classification algorithms are explored and evaluated on EEG epochs from 11 subjects in an inter-subject cross-validation setting. Due to large number of features feature ranking and selection is performed prior to classification using the ReliefF ranking algorithm within two different voting strategies. The classification models using feature subsets, achieved higher accuracy compared to the models using all features reaching 95% (Bayesian Network), 89% (Random Committee) and 87% (Random Forest) for binary classification (epileptic versus non-epileptic). The results demonstrate the competitiveness of this approach as opposed to previous methods.

Classification of epileptic and non-epileptic EEG events

In this paper, the classification of epileptic and non-epileptic events from multi-channel EEG data is investigated using a large number of time and frequency domain features. In contrast to most of the evaluations found in the literature, in this paper the non-epileptic class consists of two types of paroxysmal episodes of loss of consciousness namely the psychogenic non epileptic seizure (PNES) and the vasovagal syncope (VVS). For the classification, several classification algorithms were explored. The classification models were evaluated on EEG epochs from 11 subjects in an inter-subject cross-validation setting and the best among them achieved classification accuracies of 86% (Bayesian Network), 83% (Random Committee) and 74% (Random Forest).

Automatic estimation of the optimal AR order for epilepsy analysis using EEG signals

In this paper, we propose a computationally efficient method to estimate the optimal order of the autoregressive (AR) modeling of electroencephalographic (EEG) signals in order to use the AR coefficients as features for the analysis of EEG signals and the automatic detection of epileptic seizures. The estimation of the optimal AR-order is made using regression analysis of statistical features extracted from the samples of the EEG signals. The proposed method was evaluated in both background and ictal EEG segments using recordings from 10 epileptic patients. The experimental evaluation showed that the mean absolute error of the estimated optimal AR order is approximately 4 units.

Investigation of Sensor Placement for Accurate Fall Detection

Fall detection is typically based on temporal and spectral analysis of multi-dimensional signals acquired from wearable sensors such as tri-axial accelerometers and gyroscopes which are attached at several parts of the human body. Our aim is to investigate the location where such wearable sensors should be placed in order to optimize the discrimination of falls from other Activities of Daily Living (ADLs). To this end, we perform feature extraction and classification based on data acquired from a single sensor unit placed on a specific body part each time. The investigated sensor locations include the head, chest, waist, wrist, thigh and ankle. Evaluation of several classification algorithms reveals the waist and the thigh as the optimal locations.

EEG-based Classification of Epileptic and Non-Epileptic Events using Multi-Array Decomposition

In this paper, the classification of epileptic and non-epileptic events from EEG is investigated based on temporal and spectral analysis and two different schemes for the formulation of the training set. Although matrix representation which treats features as concatenated vectors allows capturing dependencies across channels, it leads to significant increase of feature vector dimensionality and lacks a means of modeling dependencies between features. Thus, the authors compare the commonly used matrix representation with a tensor-based scheme. TUCKER decomposition is applied to learn the essence of original, high-dimensional domain of feature space. In contrast to other relevant studies, the authors extend the non-epileptic class to both psychogenic non-epileptic seizure and vasovagal syncope. The classification schemes were evaluated on EEG epochs from 11 subjects. The proposed tensor scheme achieved an accuracy of 97,7% which is better compared to the spatiotemporal model even after trying to improve the latter by dimensionality reduction through principal component analysis and feature selection.

One-class classification of temporal EEG patterns for K-complex extraction

The purpose of this study was to detect one of the constituent brain waveforms in electroencephalography (EEG), the K-complex (KC). The role and significance of the KC include its engagement in information processing, sleep protection, and memory consolidation [1]. The method applies a two-step methodology in which first all the candidate KC waves are extracted based on fundamental morphological features imitating visual criteria. Subsequently each candidate wave is classified as KC or outlier according to its similarity to a set of different patterns (clusters) of annotated KCs. The different clusters are constructed by applying graph partitioning on the training set based on spectral clustering and exhibit temporal similarities in both signal and frequency content. The method was applied in whole-night sleep activity recorded using multiple EEG electrodes. Cross-validation was performed against visual scoring of singular generalized KCs during all sleep cycles and showed high sensitivity in KC detection.

Data fusion for paroxysmal events’ classification from EEG

BACKGROUNDnSpatiotemporal analysis of electroencephalography is commonly used for classification of events since it allows capturing dependencies across channels. The significant increase of feature vector dimensionality however introduce noise and thus it does not allow the classification models to be trained using a limited number of samples usually available in clinical studies.nnnNEW METHODnThus, we investigate the classification of epileptic and non-epileptic events based on temporal and spectral analysis through the application of three different fusion schemes for the combination of information across channels. We compare the commonly used early-integration (EI) scheme - in which features are fused from all channels prior to classification - with two late-integration (LI) schemes performing per channel classification when: (i) the temporal context varies significantly across channels, thus local spatial training models are required, and (ii) the spatial variations are negligible in comparison to the inter-subject variation, thus only the temporal variation is modeled using a single global spatial training model. Furthermore, we perform dimensionality reduction either by feature selection or by principal component analysis.nnnRESULTSnThe framework is applied on events that manifest across most channels, as generalized epileptic seizures, psychogenic non-epileptic seizures and vasovagal syncope. The three classification architectures were evaluated on EEG epochs from 11 subjects.nnnCOMPARISON WITH EXISTING METHODSnAlthough direct comparison with other studies is difficult due to the different characteristics of each dataset, the achieved recognition accuracy of the LI fusion schemes outperforms the performance reported in the literature.nnnCONCLUSIONSnThe best scheme was the LI with global model which achieved 97% accuracy.

Feature Selection Evaluation for Light Human Motion Identification in Frailty Monitoring System

In order to plan and deliver health care in a world with increasing number of older people, human motion monitoring is a must in their surveillance, since the related information is crucial for understanding their physical status. In this article, we focus on the physiological function and motor performance thus we present a light human motion identification scheme together with preliminary evaluation results, which will be further exploited within the FrailSafe Project. For this purpose, a large number of time and frequency domain features extracted from the sensor signals (accelerometer and gyroscope) and concatenated to a single feature vector are evaluated in a subject dependent cross-validation setting using SVMs. The mean classification accuracy reaches 96%. In a further step, feature ranking and selection is preformed prior to subject independent classification using the ReliefF ranking algorithm. The classification model using feature subsets of different size is evaluated in order to reveal the best dimensionality of the feature vector. The achieved accuracy is 97% which is a slight improvement compared to previous approaches evaluated on the same dataset. However, such an improvement can be considered significant given the fact that it is achieved with lighter processing using a smaller number of features.

Global vs local classification models for multi-sensor data fusion

The aim of this paper is to investigate feature extraction and fusion of information across a number of sensors in different spatial locations to classify temporal events. Although the common feature-level fusion allows capturing spatial dependencies across sensors, the significant increase of feature vector dimensionality does not allow learning the classification models using a small number of samples usually available in practice. In decision-level fusion on the other hand, sensor-specific classification models are trained and subsequently integrated to reach a combined decision. Recent work has shown that decision-level fusion with a global (common for all sensors) classification model, is more appropriate for generalized events that show a (weak or strong) manifestation across all sensors. Although we can hypothesize that the choice of scheme depends on the event type (generalized vs focal/local), the prior work does not provide enough evidence to guide on the choice of fusion scheme. Thus in this work we aim to compare the three data fusion schemes for classification of generalized and non-generalized events using two case scenarios: (i) classification of paroxysmal events based on EEG patterns and (ii) classification of falls and activities of daily living (ADLs) from multiple sensors. The results support our hypothesis that feature level fusion is more beneficial for the characterization of heterogeneous data (based on an adequate number of samples), while sensor-independent classifiers should be selected in the case of generalized manifestation patterns.

Improving Human Motion Identification Using Motion Dependent Classification

In this article, we present a new methodology for human motion identification based on motion dependent binary classifiers that afterwards fuse their decisions to identify an Activity of Daily Living (ADL). Temporal and spectral features extracted from the sensor signals (accelerometer and gyroscope) and concatenated to a single feature vector are used to train motion dependent binary classification models. Each individual model is capable to recognize one motion versus all the others. Afterwards the decisions are combined by a fusion function using as weights the sensitivity values derived from the evaluation of each motion dependent classifier on the provided training set. The proposed methodology was evaluated using SVMs for the motion dependent classifiers and is compared against the common multiclass classification approach optimized using either feature selection or subject dependent classification. The classification accuracy of the proposed methodology reaches 99% offering competitive performance comparing to the other approaches.