Electronic Theses and Dissertation

Autism spectrum disorder (asd) merupakan gangguan perkembangan saraf yang deteksi dininya masih bergantung pada observasi manual tenaga ahli terbatas, sehingga diperlukan pendekatan objektif berbasis sinyal electroencephalography (eeg). penelitian ini mengembangkan sistem deteksi dini asd dari sinyal eeg menggunakan dekomposisi ensemble empirical mode decomposition (eemd), ekstraksi fitur kompleksitas higuchi fractal dimension dan hjorth parameters, serta algoritma granular ball theory support vector machine (gbt-svm), yang diimplementasikan pada aplikasi mobile android. data terdiri dari 18 subjek (9 asd, 9 normal) yang menghasilkan 8.079 segmen dengan 192 fitur per segmen. model dilatih menggunakan parameter hasil grid search dan dievaluasi dengan 9-fold stratified cross-validation, menghasilkan akurasi rata-rata 98,66% dan f1-score 98,66%, dengan sensitivitas, spesifisitas, dan roc-auc yang turut menunjukkan kinerja tinggi dan konsisten antar fold, signifikan secara statistik dibandingkan tebakan acak. interpretasi shap menunjukkan channel ch10, komponen imf2, dan fitur mobility sebagai kontributor paling dominan. model berhasil diintegrasikan ke aplikasi mobile berarsitektur flutter, flask, dan firebase firestore, yang memproses sinyal eeg, menampilkan progres real-time, dan menyajikan hasil diagnosis beserta interpretasinya. pengujian fungsional berjalan sesuai spesifikasi, dan pengujian usability menghasilkan skor rata-rata pada kategori good hingga excellent. penelitian ini membuktikan kombinasi eemd, fitur kompleksitas, dan gbt-svm mampu menghasilkan sistem deteksi dini asd yang akurat, interpretatif, dan siap digunakan praktis melalui smartphone.

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder whose early detection still relies on manual observation by limited specialists, requiring an objective approach based on Electroencephalography (EEG) signals. This study develops an early ASD detection system from EEG signals using Ensemble Empirical Mode Decomposition (EEMD), complexity feature extraction through Higuchi Fractal Dimension and Hjorth Parameters, and the Granular Ball Theory Support Vector Machine (GBT-SVM) classification algorithm, implemented on an Android mobile application. The data consisted of 18 subjects (9 ASD, 9 Normal) producing 8,079 segments with 192 features per segment. The model was trained using grid search parameters and evaluated with 9-Fold Stratified Cross-Validation, achieving an average accuracy of 98.66% and F1-Score of 98.66%, with sensitivity, specificity, and ROC-AUC also showing consistently high performance across folds, statistically significant compared to random chance. SHAP interpretation identified channel Ch10, IMF2 component, and the Mobility feature as the most dominant contributors. The trained model was successfully integrated into a mobile application built with Flutter, Flask, and Firebase Firestore, which processes uploaded EEG signals, displays real-time progress, and presents diagnosis results with their interpretation. Functional testing showed the system operated according to specification, and usability testing produced an average score within the Good to Excellent category. This study demonstrates that combining EEMD, complexity features, and GBT-SVM produces an accurate, interpretable early ASD detection system ready for practical use through smartphone devices.