Basic Research
An Automatic Patient-Specific Seizure Onset Detection Method Using Intracranial Electroencephalography

https://doi.org/10.1111/ner.12214Get rights and content

Objective

This study presents a multichannel patient-specific seizure detection method based on the empirical mode decomposition (EMD) and support vector machine (SVM) classifier.

Materials and Methods

The EMD is used to extract features from intracranial electroencephalography (EEG). A machine-learning algorithm is used as a classifier to discriminate between seizure and nonseizure intracranial EEG epochs. A postprocessing algorithm is proposed to reject artifacts and increase the robustness of the method. The proposed method was evaluated using 463 hours of intracranial EEG recordings from 17 patients with a total of 51 seizures in the Freiburg EEG database.

Results

The proposed method had better performance than most of the existing seizure detection systems, including an average sensitivity of 92%, false detection rate (FDR) of 0.17/hour, and time delay (TD) of 12 sec. Moreover, the FDR could be further reduced by a TD extension.

Conclusions

Given its high sensitivity and low FDR, the proposed patient-specific seizure detection method can greatly assist clinical staff with automatically marking seizures in long-term EEG or detecting seizure onset online with high performance. Early and accurate seizure detection using this method may serve as a practical tool for planning epilepsy interventions.

Section snippets

INTRODUCTION

Epilepsy is a disease characterized by recurrent episodes of dysfunctional brain activity associated in time with changes in behavior. Such episodes are called seizures, and their clinical manifestations include loss of awareness or consciousness and disturbances of movement and sensation. The prevalence of epilepsy is high, affecting about 1% of the world’s population (1).

Electroencephalography (EEG) is currently one of the most important diagnostic tools for epilepsy in which seizures usually

METHODS

We treat seizure detection as a binary classification problem that involves separating seizure activity from nonseizure activity. We adopt a patient-specific approach to seizure detection to overcome the cross-patient variability and exploit the consistency within ictal patterns that emerge from the same brain region within the same patient. The proposed method consists of preprocessing, feature extraction based on the EMD technique, SVM classification, and postprocessing, and each step is

RESULTS

The proposed method was evaluated using 463 hours of iEEG recordings from 17 patients with a total of 51 seizures in the Freiburg EEG database. All of the seizures were more than 10 sec with sustained epileptic activity in several channels. The performance is evaluated by three commonly used measures of sensitivity (proportion of correctly detected seizures), FDR (number per hour), and TD (time latency between seizure onset detected and seizure onset marked by the epileptologist). The method

DISCUSSION

In this paper, we have presented a novel patient-specific approach to detecting seizures based on EMD, which extracted features of iEEG by decomposing time series into a finite and often small number of IMFs and a statistical machine learning algorithm (SVM) to discriminate between seizure and nonseizure activity. An average detection sensitivity of 92%, average FDR of 0.17/h, and average TD of 12 sec were obtained using our method. Given its high sensitivity and low FDR, this system may serve

CONCLUSIONS

Our novel method for seizure detection, which is based on EMD and SVM to assist in the binary classification of seizure and nonseizure activity, achieved better performance results than those of the existing methods. Its high sensitivity and low FDR suggest that it will be of great value for detecting epileptic seizure in iEEG recordings. Further improvements in this proposed method will aim to establish an online seizure detection system linked to therapeutic interventions in the form of a

Acknowledgements

We would like to thank Jian-min Zhang for her editorial support during the preparation of this manuscript. This work was supported by National High Technology Research and Development Program of China (No. 2012AA020408), the National Natural Science Foundation of China (No. 81301181) and the Zhejiang Provincial Science and Technology Project (No. 2013C03045-3).

Authorship Statements

Yu-xin Zheng, Jun-ming Zhu, and Yu Qi designed and conducted the study, including patient recruitment, data collection, and data

REFERENCES (20)

There are more references available in the full text version of this article.

Cited by (0)

For more information on author guidelines, an explanation of our peer review process, and conflict of interest informed consent policies, please go to http://www.wiley.com/bw/submit.asp?ref=1094-7159&site=1

1

Co-first authors.

View full text