AI Summary & Key Takeaways
AI InsightThese findings highlight the effectiveness of integrating Linear Predictive Coding and Discrete Wavelet Transform features in a lightweight and interpretable ensemble architecture, offering a promising solution for accurate and scalable seizure detection in clinical and portable settings.
📌 3 Key Takeaways
Abstract
Automated detection of epileptic seizures from electroencephalogram signals plays a critical role in clinical decision support and continuous patient monitoring. In this study, a novel hybrid feature extraction method is proposed, combining Linear Predictive Coding coefficients and statistical descriptors from Discrete Wavelet Transform sub-bands to enhance the representational capacity of EEG signals. Unlike conventional approaches that rely solely on spectral, time-domain, or entropy-based features, this method captures both temporal dynamics and localized frequency characteristics of the signal. The system was developed and evaluated using the publicly available Bonn EEG dataset, which includes both healthy and epileptic recordings in controlled conditions. Each 4096-sample EEG segment was split into two equal parts to increase data volume, resulting in 1000 segments for analysis. Feature vectors were classified using an ensemble model based on majority voting across four classifiers: Random Forest, Support Vector Machine, Multilayer Perceptron and k-Nearest Neighbors. Performance was assessed across 16 binary and multi-class classification tasks using accuracy and Matthews Correlation Coefficient as evaluation metrics. The proposed hybrid approach consistently outperformed individual feature types in all tasks, achieving up to 100% accuracy and perfect Matthews Correlation Coefficient in seizure vs. non-seizure classifications. These findings highlight the effectiveness of integrating Linear Predictive Coding and Discrete Wavelet Transform features in a lightweight and interpretable ensemble architecture, offering a promising solution for accurate and scalable seizure detection in clinical and portable settings.Automated detection of epileptic seizures from electroencephalogram signals plays a critical role in clinical decision support and continuous patient monitoring. In this study, a novel hybrid feature extraction method is proposed, combining Linear Predictive Coding coefficients and statistical descriptors from Discrete Wavelet Transform sub-bands to enhance the representational capacity of EEG signals. Unlike conventional approaches that rely solely on spectral, time-domain, or entropy-based features, this method captures both temporal dynamics and localized frequency characteristics of the signal. The system was developed and evaluated using the publicly available Bonn EEG dataset, which includes both healthy and epileptic recordings in controlled conditions. Each 4096-sample EEG segment was split into two equal parts to increase data volume, resulting in 1000 segments for analysis. Feature vectors were classified using an ensemble model based on majority voting across four classifiers: Random Forest, Support Vector Machine, Multilayer Perceptron and k-Nearest Neighbors. Performance was assessed across 16 binary and multi-class classification tasks using accuracy and Matthews Correlation Coefficient as evaluation metrics. The proposed hybrid approach consistently outperformed individual feature types in all tasks, achieving up to 100% accuracy and perfect Matthews Correlation Coefficient in seizure vs. non-seizure classifications. These findings highlight the effectiveness of integrating Linear Predictive Coding and Discrete Wavelet Transform features in a lightweight and interpretable ensemble architecture, offering a promising solution for accurate and scalable seizure detection in clinical and portable settings.
Özet
Epileptik nöbetlerin EEG sinyallerinden otomatik olarak tespit edilmesi, klinik karar desteği ve sürekli hasta takibi açısından kritik bir rol oynamaktadır. Bu çalışmada, EEG sinyallerinin temsil gücünü artırmak amacıyla, Doğrusal Öngörümleme Kodlama katsayıları ile Ayrık Dalgacık Dönüşümü alt bantlarından elde edilen istatistiksel tanımlayıcıları birleştiren yeni bir hibrit öznitelik çıkarım yöntemi önerilmiştir. Geleneksel yaklaşımların yalnızca spektral, zaman domeni veya entropi tabanlı özniteliklere dayanmasının aksine, bu yöntem hem zamansal dinamikleri hem de sinyalin yerel frekans özelliklerini yakalamaktadır. Sistem hem sağlıklı hem de epileptik kayıtları içeren ve kontrollü koşullarda elde edilmiş herkese açık Bonn EEG veri seti kullanılarak geliştirilmiş ve değerlendirilmiştir. Her biri 4096 örnekten oluşan EEG sinyalleri, veri hacmini artırmak amacıyla iki eşit parçaya bölünmüş ve analiz için toplam 1000 segmente ulaşılmıştır. Öznitelik vektörleri, Rastgele Orman, Destek Vektör Makinesi, Çok Katmanlı Algılayıcı ve k-En Yakın Komşu olmak üzere dört sınıflandırıcıdan oluşan ve çoğunluk oylamasına dayanan bir topluluk modeli ile sınıflandırılmıştır. Performans değerlendirmesi, doğruluk ve Matthews Korelasyon Katsayısı ölçütleri kullanılarak 16 adet ikili ve çok sınıflı sınıflandırma görevinde gerçekleştirilmiştir. Önerilen hibrit yaklaşım, tüm görevlerde bireysel öznitelik türlerine göre tutarlı şekilde üstün performans göstermiş; nöbet ve nöbet dışı sınıflandırmalarında %100’e kadar doğruluk ve kusursuz Matthews Korelasyon Katsayısı elde etmiştir. Bu bulgular, Doğrusal Öngörümleme Kodlama ve Ayrık Dalgacık Dönüşümü özniteliklerinin yükte hafif ve yorumlanabilir bir topluluk mimarisi içinde entegrasyonunun etkinliğini ortaya koymakta; klinik ve taşınabilir uygulamalarda doğru ve ölçeklenebilir nöbet tespiti için umut verici bir çözüm sunmaktadır.