Electronic Theses and Dissertation

Autism spectrum disorder (asd) merupakan gangguan perkembangan saraf yang diagnosisnya masih bergantung pada asesmen klinis berbasis perilaku. penelitian ini mengembangkan pipeline klasifikasi asd berbasis electroencephalography (eeg) yang akurat, interpretabel, dan efisien dalam penggunaan channel. dataset eeg king abdulaziz university yang digunakan terdiri atas 16 subjek, yaitu delapan subjek asd dan delapan subjek neurotipikal, dengan 16 channel dan frekuensi sampling 256 hz. prapemrosesan meliputi common average reference, bandpass filtering 1–40 hz, winsorizing, dan segmentasi empat detik dengan tumpang tindih 50%. setiap epoch dikonversi menjadi representasi waktu-frekuensi menggunakan continuous wavelet transform morlet dan diklasifikasikan menggunakan lightweight cnn. evaluasi dilakukan melalui 5-fold subject-wise cross-validation, sehingga seluruh epoch dari subjek yang sama hanya ditempatkan pada data latih atau data uji dalam setiap fold. model menghasilkan akurasi 95,57 ± 0,88%, f1-score 95,76 ± 0,81%, sensitivitas 96,27 ± 1,20%, spesifisitas 94,82 ± 2,11%, dan auc 0,9905 ± 0,0028. analisis shap mengidentifikasi oz, fp2, f4, dan o1 sebagai channel paling diskriminatif. konfigurasi minimum empat channel mempertahankan auc sebesar 0,9717 ± 0,0023. pada representasi cwt yang sama, lightweight cnn mengungguli svm berbasis pca yang memperoleh akurasi 84,39% dan auc 0,9184. hasil ini menunjukkan bahwa pipeline yang dikembangkan berpotensi mendukung klasifikasi eeg asd yang interpretabel dan lebih efisien, tetapi tidak ditujukan untuk menggantikan diagnosis klinis.

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition whose diagnosis still relies primarily on behavioral and clinical assessments. This study developed an electroencephalography-based ASD classification pipeline that emphasizes predictive performance, interpretability, and channel efficiency. The King Abdulaziz University EEG dataset comprised 16 participants, including eight individuals with ASD and eight neurotypical controls, recorded from 16 channels at a sampling frequency of 256 Hz. Preprocessing involved common average referencing, 1–40 Hz bandpass filtering, winsorization, and four-second segmentation with 50% overlap. Each EEG epoch was transformed into a time-frequency representation using the Morlet Continuous Wavelet Transform and classified using a Lightweight CNN. Model evaluation employed five-fold subject-wise cross-validation, ensuring that all epochs belonging to the same participant were assigned exclusively to either the training or testing set within each fold. The model achieved an accuracy of 95.57 ± 0.88%, an F1-score of 95.76 ± 0.81%, a sensitivity of 96.27 ± 1.20%, a specificity of 94.82 ± 2.11%, and an AUC of 0.9905 ± 0.0028. SHAP analysis identified Oz, Fp2, F4, and O1 as the most discriminative EEG channels. A minimum four-channel configuration retained an AUC of 0.9717 ± 0.0023. Using the same CWT representation, the Lightweight CNN outperformed the PCA-based SVM baseline, which achieved an accuracy of 84.39% and an AUC of 0.9184. These findings indicate that the proposed pipeline provides an accurate and interpretable approach to EEG-based ASD classification while reducing the required number of electrodes. The system is intended as a computational decision-support tool rather than a replacement for clinical diagnosis.