A method for detecting high-frequency oscillations using semi-supervised k-means and mean shift clustering

Abstract

This paper proposes a method to detect the high-frequency oscillations (HFOs) in epileptic seizure onset zones (SOZs) localization using semi-supervised k-means and mean shift algorithm. Wavelet entropy (WE) and teager energy operator (TEO) are adopted to distinguish HFOs from normal electroencephalogram (EEG). Labeled data are used to initialize the clustering center of semi-supervised k-means algorithm, and unlabeled data are employed to obtain physiological and suspected pathological HFOs. For the suspected pathological HFOs, the mean shift algorithm is used for clustering, and the results are analyzed by the spectral center algorithm to locate SOZs. By comparing the EEG data of five patients with the results of the other three methods, it can be seen that the method proposed in this paper has good sensitivity and specificity, which is helpful for accurate localization before clinical epilepsy surgery.

Introduction

Epilepsy is a worldwide epidemic of non-communicable chronic brain disease affecting all age groups. The prevalence of epilepsy accounts for about 1% of the world's population. At present, there are about 50 million people with epilepsy worldwide, and 80% of them are in developing countries. 70% of newly diagnosed epilepsy patients can be cured with antiepileptic drugs. For patients with poor drug treatment, surgery is another possible effective way to treat epilepsy [1], [2], [3]. Preoperative accurate positioning of epileptic seizures is more important than surgery itself. The traditional seizure onset zones (SOZs) are focused on electroencephalogram (EEG) data with a frequency less than 80 Hz. The epileptic origins can be located by extracting the frequency components of EEG signals [4], [5]. Since the epileptiform discharges have a wide range of transmission, and epilepsy brain telecoms in low frequency bands are easily interfered by other signals, the misjudgment of results broadly exist [6]. From the 1990s, scholars have begun to study high-frequency oscillations (HFOs) with frequencies above 80 Hz in epileptic EEG signals [7]. Researchers classify HFOs into three types according to their frequency: Rs (Ripples, 80–250 Hz), FRs (Fast Ripples, 250–500 Hz), VHFO (Very High Frequency Oscillatory, 1000–2500 Hz) [8], [9]. Among them, VHFO are difficult to be collected and their mechanism of action is not clear. A large number of literatures are related to the Rs and FRs of HFOs in EEG data and show that Rs and FRs more directly reflect the synchronization activities of neurons. Therefore, HFOs of 80–500 Hz become a new biological indicator in the location of SOZs. It can indicates the epileptic seizure area more accurately than the epileptic discharge, which can reduce the pain caused by long-term EEG recording and reduce the possibility of surgical infection [10], [11].

Current SOZs location methods can be classified as visual assessment and automated detection. Visual assessment is the most commonly used at present [12], and HFOs are visually labeled by experts according to experience. However, artificially labeled HFOs in visual assessment are very time consuming, which will greatly increases the patient's pain and the risk of surgical infection [13]. In addition, the accuracy of the visually marking will be influenced by doctors and various objective factors. Since missing labels and mislabeling are difficult to avoid, it is necessary to find a fast and efficient automatic detection method [14]. With the in-depth research of computational intelligence methods, more computational intelligence methods have been applied in different fields. In the literature [15], a high-dimensional feature is introduced to the facial expression recognition. Furthermore, the deep sparse autoencoders (DSAE) are established to recognize the facial expressions with high accuracy by learning robust and discriminative features from the data. In the literature [16], the SDPSO-SVM model was used to classify Alzheimer's disease (AD) and mild cognitive impairment (MCI), which became an effective diagnostic method for AD.

Recent studies have shown that HFOs are closely related to SOZs, and can quickly and accurately locate SOZs by detecting and judging HFOs. In related researches, automatic detection includes feature extraction and classification of HFOs. Detection algorithms in time domain, frequency domain and time-frequency domain, including root mean square (RMS) [17], short-term linear length [18], complex Morlet wavelet transform [19] and Hilbert transform [20] are used in the feature extraction. In [21], RMS is calculated from the filtered EEG signal to estimate the short-term energy. However, that method cannot accurately predict the EEG signal and its wake signal. In [22], the complex Morlet wavelet is used to extract the time-frequency characteristics of EEG from the wavelet transform of the mother wavelet. However, the time-frequency resolution of different patient data is not high. The classification methods of HFOs mainly include threshold method, fuzzy neural network (FNN), support vector machine (SVM) classification and so on. In [23], a method for automatic detection of HFOs is proposed. The fuzzy entropy and short-time energy are extracted from the EEG signal as feature input to the FNN classifier. Through the FNN classifier, the signal can be divided into HFOs and normal rhythm. In [24], Gabor transform is performed on EEG data, and the energy ratio and duration characteristics of HFOs events are extracted. Simultaneously input and train multiple types of SVM classifiers for various types of HFOs and artificial noise. The method in [25] first detects the event of suspected HFOs, and uses the continuous wavelet transform and the improved method of short-time Fourier transform to transform the time-frequency diagram of the suspected event, and considers that HFOs are in time. The frequency map shows an independent bright spot. Moreover, it is judged whether the suspected event is a HFO according to the frequency and power spectral density. The location of the epileptic SOZ is to count the frequency of occurrence of these events. In [26], a multi-feature extraction and threshold comprehensive judgment method is proposed. That method has high sensitivity, but the steps are complicated and the calculation time is long. Based on the latest research from abroad, it is still in an exploratory research stage. Although these methods have high sensitivity, their specificity is low, and there are still false positives.

Clustering algorithms have been widely used in many aspects. For example, a new clustering algorithm is proposed in [27], [28] to improve the accuracy of traditional clustering approaches with applications in analyzing real-time patient attendance data from an accident & emergency (A&E) department in a local UK hospital.

In this paper, the semi-supervised k-means and mean shift algorithms are used to analyze and process HFOs. The data are preprocessed, then the feature is composed of feature vectors, and the labeled center is used to initialize the cluster center of the k-means algorithm. In practical application, the original EEG of each channel of the patient is preprocessed and extracted feature, and the clustered center is used for clustering. The obtained suspected pathological HFOs are clustered by the mean shift algorithm, and the results are analyzed by using the spectral centroid (SC) to determine the channel of the SOZ position. The analytical method improves the sensitivity and specificity of the algorithm, and lays a solid theoretical and experimental basis for the accurate positioning of the epileptic SOZs.