Representation based on ordinal patterns for seizure detection in EEG signals
Yunxiao Liu1, Youfang Lin1, Ziyu Jia1
1Beijing Key Lab of Traffic Data Analysis and Mining, School of Computer and Information Technology, Beijing Jiaotong University, Beijing, 100044, PR China; CAAC Key Laboratory of Intelligent Passenger Service of Civil Aviation, Beijing, China.
Computers in Biology and Medicine
|October 22, 2020
Summary
This study introduces novel unigram ordinal pattern (UniOP) and bigram ordinal pattern (BiOP) representations for analyzing electroencephalogram (EEG) signals. These methods effectively detect seizures with over 90% accuracy, outperforming existing techniques.
Area of Science:
- Neuroscience
- Signal Processing
- Machine Learning
Background:
- Electroencephalogram (EEG) signals are crucial for diagnosing epilepsy, but current seizure detection algorithms often require complex feature engineering.
- Existing methods rely heavily on domain-specific knowledge and expertise, limiting their accessibility and broad applicability.
Purpose of the Study:
- To introduce novel unigram ordinal pattern (UniOP) and bigram ordinal pattern (BiOP) representations for analyzing time series data, specifically EEG signals.
- To develop a seizure detection method that minimizes reliance on domain-specific features and expertise.
- To evaluate the effectiveness of UniOP and BiOP representations in capturing the dynamics of EEG signals during different states (healthy, seizure-free, seizure).
Main Methods:
- Transforming time series subsequences into ordinal patterns based on value rankings to create UniOP representations.
- Analyzing the co-occurrence of ordinal patterns to create BiOP representations, capturing contextual information.
- Combining UniOP and BiOP representations with a nearest neighbor algorithm for seizure detection and evaluating performance on three public EEG datasets.
Main Results:
- The proposed method achieved over 90% accuracy, sensitivity, and specificity on the Bonn EEG dataset, outperforming several state-of-the-art methods.
- Performance on a second dataset was comparable to existing state-of-the-art techniques, indicating good generalization capabilities.
- Consistent performance with approximately 89% for all metrics on a third, large-scale dataset, demonstrating suitability for extensive data.
Conclusions:
- UniOP and BiOP representations effectively capture essential dynamics in EEG time series across healthy, seizure-free, and seizure states.
- The proposed method offers a robust and accessible approach to seizure detection, reducing the need for intricate feature engineering.
- The approach demonstrates strong performance and generalization, making it suitable for various EEG analysis tasks, including large-scale applications.
