Automated seizure detection using limited-channel EEG and non-linear dimension reduction

https://doi.org/10.1016/j.compbiomed.2017.01.011Get rights and content

Abstract

Electroencephalography (EEG) is an essential component in evaluation of epilepsy. However, full-channel EEG signals recorded from 18 to 23 electrodes on the scalp is neither wearable nor computationally effective. This paper presents advantages of both channel selection and nonlinear dimension reduction for accurate automatic seizure detection. We first extract the frequency domain features from the full-channel EEG signals. Then, we use a random forest algorithm to determine which channels contribute the most in discriminating seizure from non-seizure events. Next, we apply a non-linear dimension reduction technique to capture the relationship among data elements and map them in low dimension. Finally, we apply a KNN classifier technique to discriminate between seizure and non-seizure events. The experimental results for 23 patients show that our proposed approach outperforms other techniques in terms of accuracy. It also visualizes long-term data in 2D to enhance physician cognition of occurrence and disease progression.

Introduction

Many patients gain control of their seizures with anti-convulsant or anti-epileptic drugs (AEDs); However, approximately one-third of the patients prove intractable with medications [1]. Serious physical injury and death (known as Sudden Unexpected Death of Epilepsy or SUDEP) are seen in patients with uncontrolled epilepsy. Risk of serious physical injury or even death is high for such epileptic patients. There is a need, therefore, for accurate, automated detection of seizures to alert patients/caregivers and to help epileptologist provide significant therapeutic solutions and to improve the quality of life of epileptic patients [2].

Electroencephalography (EEG) signals are the best indicators of epileptic seizures. Epileptiform EEG patterns, such as spikes and sharp waves, can assist in the diagnosis and in classifying seizures [3]. EEG signal analysis, as a diagnostic test, has had an enormous impact on epileptic seizure detection [4]. EEG signal analysis can improve our understanding of abnormal brain activities like neurons misfiring or excessive neural activity. However, EEG-based seizure detection is relatively complex since it involves voluminous amounts of recorded EEG samples coming from multiple scalp mounted electrodes [5].

In order to tackle the problems related to the complexity of EEG-based seizure detection, two techniques can be employed: First, select a limiting number of EEG channels instead of using all EEG (often 18–23) electrodes. Full-channel EEG monitoring is only practical in clinics since it is expensive, time consuming, uncomfortable and stigmatizing. Moreover, irrelevant channels add noise to the feature space and decrease the seizure detection accuracy. Limited-channel configurations can be implemented with newly developed discrete wearable patches (e.g., Epitel nodes [6]) which are able to record seizure activity in daily life using individual nodes. There are recently developed techniques which make these wearable devices more energy-efficient and practical for daily life applications [7]. Second, map the high dimensional dataset into a lower dimension by preserving the neighborhood information of data points. Since the feature space of EEG signals has a non-linear structure, non-linear data embedding (that preserves distance information among neighboring data points) techniques are the best option for reducing the number of feature space dimensions. In this work, we combine both channel selection and dimension reduction to provide a low dimensional feature space for epileptic seizure detection.

There are two different approaches to handling the problem of high-dimensionality in EEG signals:

  • (1)

    Channel selection: These techniques select an effective subset of the original channels. A filter-based channel selection technique based on maximum variance criteria has been investigated in [8]. Authors used the difference in variance between seizure and non-seizure events for selecting the best number of channels in [9]. The information gain of EEG channels is used in [10] to find the most informative channels for seizure detection. Authors calculated the accuracy of all individual EEG channels and selected the channels with the highest accuracy for limited channel seizure detection in [11]. However, these techniques may not always be efficient, particularly when there are a huge number of features extracted from a noisy environment like scalp EEG.

  • (2)

    Dimension reduction: Researchers have investigated how to map an original feature space into a smaller feature space representation in order to reduce the overwhelming number of features. Authors in [12] presented a semi-automated patient-dependent unsupervised technique using all EEG channels and Principal Component Analysis (PCA) as their dimension reduction technique. Authors in [13] applied autocorrelation to extract features and used Common Spatial Patterns (CSP) to decrease the feature space dimensionality. Authors in [11] extracted 26 features per channel and reduced the number to 8 using PCA. They reported the accuracy of EEG channels individually using an LDA classifier. Although PCA is able to reduce the number of features, it cannot properly preserve the distance relationships between data points in low dimensions. The number of patients investigated in [11] is too small to be conclusive (only six epileptic patients out of 23 available patients in the MIT database [14]). Moreover, the number of non-seizure events selected is too small for adequate training.

To the best of our knowledge, our proposed method is the first work which combines channel-selection and dimension reduction for EEG-based epileptic seizure detection. The general view of our proposed epileptic seizure detection model is shown in Fig. 1. First, we use power spectral analysis to extract features per channel per subject for each time window. Then, we use a random decision forest for selecting a limited number of channels. To do this, we first generate a large number of random decision trees using features from all the EEG channels. Next, we investigate the number of times each channel appears in the forest. The channels with the highest contribution to the entire forest are chosen for limited channel seizure detection. The best channels are selected by voting among all the decision trees in the forest, thereby making our method robust against noisy channels - the most important advantage of this technique. Next, we use t-distributed stochastic neighbor embedding (t-SNE) to embed and represent data in a lower dimension by preserving the relationship of data points in high dimensional feature space. Finally, we use a KNN classifier model to differentiate between seizure and non-seizure events in a patient-specific manner.

Section snippets

Feature extraction

Feature Extraction is a critical step in EEG-based seizure detection since it extracts seizure related characteristics. Researchers categorize EEG signals into unique bandwidths as shown in Table 1. Both normal and abnormal brain functions are listed by bandwidth [15]. Note that seizures affect most of the EEG frequencies. Consequently, frequency features are widely used for EEG-based epileptic seizure detection [16].

In our method, we first segmented the EEG signals into 10 s windows. Next, we

Motivation & challenges

Limited-channel EEG epileptic seizure detection offers three advantages. First, it reduces the computational complexity of seizure detection, leading to a faster run time and lower power consumption, thereby making seizure detection models faster and more cost-effective (e.g. for inexpensive embedded systems). Second, for some patients, it increases the detection accuracy by avoiding redundancy of non-focal/unnecessary channels. Finally, reduction from 23 to 1–3 channels makes wearable EEG

Dimension reduction

Using our technique, the total number of channels is significantly reduced (e.g. k=3 vs. originally 18–23 channels). Consequently, the number of spectral features is also reduced (e.g. 15 vs. originally 90–115 features). However, 15 is still a high dimensional feature space suffering from the curse of dimensionality. The main problem is the empty space phenomenon, which means data points become more and more sparse as the feature space dimension increases. This phenomenon does not allow

Classification

For each patient, we selected 8 h of known nonseizure data that was at least 30 min from the nearest seizure. For this group of patients, seizure duration varies from 10 to 120 s. We segmented EEG signals into 10 s windows to accommodate the shortest seizures. We refer to data points (instances) as segmented windows, which are used for spectral feature extraction as discussed in Section 2. After channel selection and dimension reduction, each data point is represented by only two features (2D

EEG Dataset

Twenty three epileptic patients' EEG manifestations from the open access scalp EEG database is used in this work. The dataset was recorded from twenty three pediatric patients with intractable seizures at the Children's Hospital Boston [33]. The dataset is available at the PhysioNet website: www.physionet.org/pn6/chbmit [14]. Data collection was done while patients were off anti-seizure medications to capture EEG responses to seizure and thereby determine which patients were candidates for

Conclusion

Full channel EEG analysis constrains the performance of seizure detection since it provides extremely high dimensional feature spaces. Confounding effects of irrelevant EEG channels can decrease the performance of seizure detection techniques. Generating a large number of random decision trees using features coming from all the EEG channels is an efficient way to select a small number of EEG channels that can most effectively detect the patient's seizures. Relevant and informative features

References (36)

  • R. Saeedi, R. Fallahzadeh, P. Alinia, H. Ghasemzadeh, An energy-efficient computational model for uncertainty...
  • M. Qaraqe, M. Ismail, Q. Abbasi, E. Serpedin, Channel selection and feature enhancement for improved epileptic seizure...
  • D. Cogan, J. Birjandtalab, M. Nourani, J. Harvey, V. Nagaraddi, Multi-biosignal analysis for epileptic seizure...
  • M. Yıldız, E. Bergil, The investigation of channel selection effects on epileptic analysis of eeg...
  • O. Smart, M. Chen, Semi-automated patient-specific scalp eeg seizure detection with unsupervised machine learning, in:...
  • S. Khanmohammadi, C.-A. Chou, A simple distance based seizure onset detection algorithm using common spatial patterns,...
  • A.L. Goldberger et al.

    Physiobank, physiotoolkit, and physionet components of a new research resource for complex physiologic signals

    Circulation

    (2000)
  • W.O. Tatum

    Ellen r. grass lecture: extraordinary eeg

    Neurodiagnostic J.

    (2014)
  • Cited by (117)

    • A self-attention model for cross-subject seizure detection

      2023, Computers in Biology and Medicine
    View all citing articles on Scopus
    View full text