Electronic Theses and Dissertation

Klasifikasi Autism Spectrum Disorder (ASD) berbasis Electroencephalography (EEG) umumnya dievaluasi pada level segmen sinyal tanpa validasi independen antar subjek, sehingga rentan terhadap kebocoran data dan estimasi performa yang terlalu optimistis. Penelitian ini mengusulkan kerangka kerja agregasi fitur tingkat-subjek (subject-level aggregation) berbasis fitur Wavelet Packet Transform (WPT) untuk klasifikasi ASD berbasis sinyal EEG 16-kanal. Data dikumpulkan dari 16 partisipan (8 ASD, 8 neurotipikal) di Banda Aceh, Indonesia, menggunakan OpenBCI Cyton 16-kanal pada frekuensi sampling 250 Hz. Sinyal dipra-pemrosesan melalui notch filtering dan Butterworth bandpass filtering (0,5–40 Hz), kemudian tiga strategi penghilangan artefak dibandingkan pada konfigurasi downstream yang identik: Infomax ICA sebagai baseline, Multiscale ICA-DWT (MSICA-DWT), dan Learnable Wavelet Packet Transform berbasis ICA (LWPT-ICA-iLWPT). Sinyal disegmentasi menjadi jendela 4 detik (overlap 50%) dan diekstraksi menjadi fitur energi relatif dan varians WPT, diseleksi menggunakan ANOVA F-value secara fold-wise, diagregasi ke tingkat subjek (mean, standar deviasi, median, IQR), lalu diklasifikasikan menggunakan Logistic Regression beregularisasi L2. Evaluasi dilakukan melalui nested Leave-One-Subject-Out Cross-Validation (LOSO-CV), validasi subject-pair berulang (50 kali), dan permutation test. Pipeline dengan Infomax ICA mencapai balanced accuracy 0,9375 (F1-score 0,941) pada nested LOSO-CV, serta balanced accuracy 0,9525 ± 0,0529 dan AUC 0,9897 ± 0,0163 pada evaluasi subject-pair berulang, dengan signifikansi yang dikonfirmasi permutation test (p = 0,0196). Wilcoxon Signed-Rank Test menunjukkan Infomax ICA secara signifikan mengungguli MSICA-DWT pada keenam metrik evaluasi (p < 0,001; effect size r = 1,00). Hasil ini menunjukkan bahwa strategi penghilangan artefak konvensional berbasis ICA, dikombinasikan dengan agregasi fitur tingkat subjek dan validasi subject-independent yang ketat, menghasilkan pipeline klasifikasi ASD berbasis EEG yang akurat dan relevan secara klinis.

Kata Kunci: Autism Spectrum Disorder (ASD), Electroencephalography (EEG), Wavelet Packet Transform, agregasi tingkat subjek, Infomax ICA, Logistic Regression, nested LOSO cross-validation.

EEG-based classification of Autism Spectrum Disorder (ASD) is commonly evaluated at the signal-segment level without subject-independent validation, making it prone to data leakage and overly optimistic performance estimates. This study proposes a subject-level feature aggregation framework based on Wavelet Packet Transform (WPT) energy and variance features for ASD classification using 16-channel EEG signals. Data were collected from 16 participants (8 ASD, 8 neurotypical) in Banda Aceh, Indonesia, using an OpenBCI Cyton system at a 250 Hz sampling rate. Signals were preprocessed through notch filtering and Butterworth bandpass filtering (0.5–40 Hz), after which three artifact-removal strategies were compared under an identical downstream configuration: Infomax ICA as the baseline, Multiscale ICA-DWT (MSICA-DWT), and ICA-based Learnable Wavelet Packet Transform (LWPT-ICA-iLWPT). Signals were segmented into 4-second windows (50% overlap) and decomposed into WPT relative-energy and variance features, selected using the ANOVA F-value in a fold-wise manner, aggregated to the subject level (mean, standard deviation, median, IQR), and classified using L2-regularized Logistic Regression. Evaluation was performed through nested Leave-One-Subject-Out Cross-Validation (LOSO-CV), repeated subject-pair validation (50 iterations), and a permutation test. The Infomax ICA pipeline achieved a balanced accuracy of 0.9375 (F1-score 0.941) under nested LOSO-CV, and a balanced accuracy of 0.9525 ± 0.0529 with an AUC of 0.9897 ± 0.0163 under repeated subject-pair evaluation, with significance confirmed by a permutation test (p = 0.0196). A Wilcoxon Signed-Rank Test showed that Infomax ICA significantly outperformed MSICA-DWT across all six evaluation metrics (p < 0.001; effect size r = 1.00). These results indicate that a conventional ICA-based artifact-removal strategy, combined with subject-level feature aggregation and strict subject-independent validation, yields an EEG-based ASD classification pipeline that is accurate and clinically relevant. Keywords: Autism Spectrum Disorder (ASD), Electroencephalography (EEG), Wavelet Packet Transform, subject-level aggregation, Infomax ICA, Logistic Regression, nested LOSO cross-validation.