Electronic Theses and Dissertation

Autism Spectrum Disorder (ASD) merupakan gangguan perkembangan saraf yang umumnya diidentifikasi melalui asesmen perilaku klinis. Meskipun tetap menjadi acuan utama, asesmen tersebut membutuhkan tenaga ahli dan waktu pemeriksaan, sehingga pendekatan komputasional berbasis biomarker neurofisiologis diperlukan sebagai sistem pendukung. Penelitian ini mengembangkan dan mengevaluasi kerangka Hybrid Graph-Spectral Feature Learning untuk klasifikasi ASD berbasis electroencephalography (EEG). Evaluasi dilakukan pada Dataset Banda Aceh dan Dataset KAU, yang masing-masing terdiri atas 7 subjek ASD dan 7 subjek kontrol dengan frekuensi sampling 250 Hz dan 256 Hz. Tahapan prapemrosesan meliputi bandpass filter 0,5–45 Hz, notch filter spesifik dataset, common average reference, dan segmentasi menjadi epoch 30 detik tanpa overlap. Delapan strategi fitur yang dibandingkan adalah NVG, HVG, PSD, rPSD, NVG_PSD, HVG_PSD, NVG_rPSD, dan HVG_rPSD. Seluruh strategi dikombinasikan dengan Support Vector Machine, XGBoost, Multilayer Perceptron, dan InceptionTime sehingga menghasilkan 64 kombinasi eksperimen. Evaluasi menggunakan 7-fold subject-wise balanced cross-validation, dengan seluruh epoch dari subjek uji dipisahkan sepenuhnya dari data pelatihan. Performa dinilai menggunakan Accuracy, Precision, Sensitivity, Specificity, dan F1-score, sedangkan perbedaan performa global dianalisis menggunakan uji Friedman. Pada Dataset Banda Aceh, kombinasi NVG_PSD dengan SVM memberikan hasil terbaik dengan Accuracy 97,37 ± 2,95% dan F1-score 97,19 ± 3,20%. Pada Dataset KAU, kombinasi rPSD dengan SVM mencapai Accuracy 97,79 ± 2,54% dan F1-score 97,66 ± 2,82%. Hasil ini menunjukkan bahwa representasi graph-spectral berbasis PSD efektif pada Dataset Banda Aceh, sedangkan fitur spektral relatif memberikan performa terbaik pada Dataset KAU. Temuan tersebut menegaskan bahwa efektivitas representasi fitur dipengaruhi oleh karakteristik dataset, sementara validasi berbasis subjek memberikan evaluasi yang lebih relevan terhadap generalisasi pada subjek baru.

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition commonly identified through clinical behavioral assessment. Although behavioral assessment remains the primary clinical standard, it requires expert involvement and considerable assessment time, motivating the development of neurophysiological biomarker-based computational approaches as supporting tools. This study developed and evaluated a Hybrid Graph-Spectral Feature Learning framework for electroencephalography (EEG)-based ASD classification. The evaluation used the Banda Aceh and KAU datasets, each comprising 7 ASD and 7 control subjects recorded at sampling frequencies of 250 Hz and 256 Hz, respectively. Preprocessing consisted of 0.5–45 Hz bandpass filtering, dataset-specific notch filtering, common average referencing, and segmentation into non-overlapping 30-s epochs. Eight feature strategies were compared: NVG, HVG, PSD, rPSD, NVG_PSD, HVG_PSD, NVG_rPSD, and HVG_rPSD. These strategies were combined with Support Vector Machine, XGBoost, Multilayer Perceptron, and InceptionTime, resulting in 64 experimental combinations. Performance was evaluated using seven-fold subject-wise balanced cross-validation, with all epochs from each test subject completely excluded from the training data. Accuracy, Precision, Sensitivity, Specificity, and F1-score were used as evaluation metrics, while global performance differences were analyzed using the Friedman test. On the Banda Aceh dataset, NVG_PSD with SVM achieved the best performance, with an Accuracy of 97.37 ± 2.95% and an F1-score of 97.19 ± 3.20%. On the KAU dataset, rPSD with SVM achieved an Accuracy of 97.79 ± 2.54% and an F1-score of 97.66 ± 2.82%. These findings show that the PSD-based graph-spectral representation was effective for the Banda Aceh dataset, whereas relative spectral features provided the best performance for the KAU dataset. The results indicate that feature effectiveness depends on dataset characteristics, while subject-wise validation provides a more relevant assessment of generalization to unseen subjects.