Electronic Theses and Dissertation

Electroencephalography (Sinyal electroencephalography (EEG) memiliki karakteristik yang berubah terhadap waktu sehingga pemilihan representasi menjadi bagian penting dalam proses klasifikasi. Penelitian ini bertujuan membentuk dan membandingkan spektrogram Short-Time Fourier Transform (STFT) dan matriks Toeplitz berbasis autokorelasi untuk klasifikasi Autism Spectrum Disorder (ASD) dan non-ASD menggunakan EfficientNetV2-S berbasis transfer learning. Data yang digunakan berasal dari 10 subjek anak, terdiri atas lima subjek ASD dan lima subjek non-ASD, serta menghasilkan 4.490 citra pada masing-masing representasi. Perekaman dilakukan menggunakan 16 kanal EEG dengan frekuensi sampling 250 Hz dan durasi 900 detik pada setiap subjek. Sinyal dipraproses menggunakan Butterworth band-pass filter 0,5–45 Hz, Independent Component Analysis, dan ICLabel, kemudian disegmentasi menjadi empat detik dengan overlap 50%. Setiap segmen diubah secara terpisah menjadi citra RGB STFT dan Toeplitz berukuran 384 × 384 piksel. Kedua representasi digunakan sebagai masukan EfficientNetV2-S dan dievaluasi menggunakan stratified five-fold cross-validation dengan konfigurasi pelatihan yang sama. Hasil pengujian menunjukkan bahwa representasi STFT menghasilkan performa lebih tinggi dengan rata-rata accuracy sebesar 96,28%, dibandingkan matriks Toeplitz yang memperoleh 92,58%. Dengan demikian, representasi STFT lebih efektif untuk klasifikasi citra EEG ASD dan non-ASD menggunakan EfficientNetV2-S pada konfigurasi penelitian ini.

Electroencephalography (EEG) signals exhibit time-varying characteristics, making the selection of an appropriate representation an important aspect of the classification process. This study aimed to construct and compare Short-Time Fourier Transform (STFT) spectrograms and autocorrelation-based Toeplitz matrices for classifying Autism Spectrum Disorder (ASD) and non-ASD using transfer learning with EfficientNetV2-S. The data were obtained from 10 children, consisting of five ASD subjects and five non-ASD subjects, and produced 4,490 images for each representation. EEG recordings were acquired using 16 channels at a sampling frequency of 250 Hz for 900 seconds per subject. The signals were preprocessed using a 0.5–45 Hz Butterworth band-pass filter, Independent Component Analysis, and ICLabel, and were subsequently segmented into four-second windows with 50% overlap. Each segment was separately transformed into 384 × 384-pixel RGB images using STFT and Toeplitz representations. Both representations were used as inputs to EfficientNetV2-S and evaluated using stratified five-fold cross-validation under identical training configurations. The experimental results showed that the STFT representation achieved higher performance, with a mean accuracy of 96.28%, compared with 92.58% obtained by the Toeplitz matrix representation. Therefore, under the experimental configuration employed, STFT was more effective for classifying ASD and non-ASD EEG images using EfficientNetV2-S.