Klasifikasi Multikelas Kejang Epilepsi Menggunakan Metode Deep Learning

Eviyanti, Ade (2026) Klasifikasi Multikelas Kejang Epilepsi Menggunakan Metode Deep Learning. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7025211029-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Epilepsi merupakan gangguan sistem saraf pusat yang ditandai oleh aktivitas listrik otak berlebihan dan dapat menyebabkan kejang berulang. Electroencephalogram (EEG) digunakan untuk merekam aktivitas tersebut, tetapi sinyal EEG bersifat kompleks, tidak stasioner, beramplitudo kecil, dan mudah terpengaruh derau. Penelitian terdahulu telah banyak mengembangkan metode deteksi kejang berbasis EEG, tetapi sebagian besar masih berfokus pada deteksi biner antara kondisi non-kejang dan kejang atau hanya membedakan tiga kelas utama. Deteksi lima kelas yang mencakup tiga tahap prakejang secara berurutan masih menghadapi masalah kemiripan pola antarkelas, tingginya jumlah channel dan fitur, serta belum optimalnya integrasi representasi waktu-frekuensi, pembelajaran pola lokal, dan hubungan temporal. Penelitian ini bertujuan mengembangkan kerangka deteksi multikelas kejang epilepsi untuk mendeteksi lima kondisi EEG secara otomatis. Penelitian menggunakan CHB-MIT Scalp EEG Database. Rekaman EEG disegmentasi menggunakan fixed-length segmentation menjadi 2.500 segmen seimbang dan tidak tumpang tindih. Setiap segmen berdurasi 10 detik dan terdiri atas 2.560 sampel pada frekuensi 256 Hz. Segmen dikelompokkan menjadi lima kelas, yaitu non-kejang, prakejang 1 pada 30 sampai 20 menit sebelum onset, prakejang 2 pada 20 sampai 10 menit sebelum onset, prakejang 3 pada 10 sampai 0 menit sebelum onset, dan kejang. Tahap pemrosesan mengintegrasikan seleksi channel berbasis koefisien variasi, dekomposisi Discrete Wavelet Transform (DWT) empat level, ekstraksi fitur menggunakan Convolutional Neural Network (CNN), perhitungan sembilan deskriptor statistik, dan seleksi fitur menggunakan Analysis of Variance (ANOVA). Fitur terpilih selanjutnya digunakan untuk mendeteksi kelas EEG menggunakan model CNN, BiLSTM, dan CNN-BiLSTM. Kontribusi penelitian meliputi pembentukan dataset EEG lima kelas dengan tiga interval prakejang yang berurutan, pengembangan rangkaian pemrosesan yang menghasilkan tujuh channel dan 50 fitur terpilih, serta integrasi CNN-BiLSTM untuk mendeteksi pola lokal dan hubungan temporal pada sinyal EEG. Evaluasi dilakukan melalui lima skenario pengujian, perbandingan model machine learning dan deep learning, studi ablasi, serta 10-fold cross-validation. Konfigurasi terbaik menggunakan tujuh channel hasil seleksi, DWT empat level, ekstraksi fitur CNN, sembilan deskriptor statistik, seleksi ANOVA, dan model CNN-BiLSTM. Pada pembagian data tunggal, model menghasilkan akurasi 98,40%, precision 98,46%, recall 98,40%, dan F1-score 98,39%. Pada 10-fold cross-validation, model memperoleh rata-rata akurasi 92,88% dan weighted F1-score 92,87%. Hasil tersebut menunjukkan bahwa integrasi reduksi channel, representasi waktu-frekuensi, pembelajaran fitur lokal, seleksi fitur statistik, dan pemodelan temporal dua arah efektif untuk mendeteksi lima tahap kejang epilepsi. Namun, validasi berbasis pasien dan pengujian pada dataset eksternal masih diperlukan untuk menilai kemampuan generalisasi model pada kondisi klinis yang lebih realistis.
========================================================================================================================================
Epilepsy is a central nervous system disorder characterized by excessive electrical activity in the brain that may cause recurrent seizures. Electroencephalography (EEG) records this electrical activity, but EEG signals are complex, nonstationary, low in amplitude, and susceptible to noise. Previous studies have developed various EEG-based seizure detection methods. However, most studies have focused on binary detection between non-seizure and seizure conditions or classification into only three major classes. Five-class detection involving three consecutive preictal stages remains challenging because of the similarity of signal patterns across classes, the large number of EEG channels and features, and the limited integration of time-frequency representations, local feature learning, and temporal dependency modeling. This study develops a multiclass epileptic seizure detection framework for automatically detecting five EEG conditions. This study used the CHB-MIT Scalp EEG Database. The EEG recordings were processed using fixed-length segmentation to produce 2,500 balanced and non-overlapping segments. Each segment had a duration of 10 seconds and contained 2,560 samples recorded at a sampling frequency of 256 Hz. The segments were categorized into five classes: non-seizure, preictal 1 at 30 to 20 minutes before seizure onset, preictal 2 at 20 to 10 minutes before seizure onset, preictal 3 at 10 to 0 minutes before seizure onset, and seizure. The processing framework integrated coefficient-of-variation-based channel selection, four-level Discrete Wavelet Transform (DWT) decomposition, feature extraction using a Convolutional Neural Network (CNN), calculation of nine statistical descriptors, and feature selection using Analysis of Variance (ANOVA). The selected features were then used to detect the EEG classes using CNN, BiLSTM, and CNN-BiLSTM models. The contributions of this study include the construction of a five-class EEG dataset with three consecutive preictal intervals, the development of a processing framework that produced seven selected channels and 50 discriminative features, and the integration of CNN-BiLSTM to detect local patterns and temporal dependencies in EEG signals. The framework was evaluated through five experimental scenarios, comparisons between machine learning and deep learning models, an ablation study, and 10-fold cross-validation. The best configuration combined seven selected channels, four-level DWT decomposition, CNN-based feature extraction, nine statistical descriptors, ANOVA-based feature selection, and CNN-BiLSTM classification. Under a single train-test split, the model achieved an accuracy of 98.40%, a precision of 98.46%, a recall of 98.40%, and an F1-score of 98.39%. Under 10-fold cross-validation, the model achieved a mean accuracy of 92.88% and a mean weighted F1-score of 92.87%. These results indicate that integrating channel reduction, time-frequency representation, local feature learning, statistical feature selection, and bidirectional temporal modeling is effective for detecting five epileptic seizure stages. However, patient-independent validation and testing on external datasets are still required to assess the model’s generalizability under more realistic clinical conditions.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	ANOVA, CHB-MIT, CNN-BiLSTM, DWT, EEG signals, epileptic seizure classification
Subjects:	Q Science > QP Physiology > Q376.5 Electroencephalography (EEG)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Ade Eviyanti
Date Deposited:	31 Jul 2026 02:32
Last Modified:	31 Jul 2026 02:32
URI:	http://repository.its.ac.id/id/eprint/141590

Actions (login required)

View Item