Electronic Theses and Dissertation

Prevalensi Autism Spectrum Disorder (ASD) yang terus meningkat dalam beberapa dekade terakhir menegaskan kebutuhan akan pendekatan skrining awal yang lebih objektif, terukur, dan berbasis data. Elektroensefalografi (EEG) merupakan modalitas neurofisiologis yang potensial karena mampu merekam dinamika aktivitas otak dengan resolusi temporal tinggi dan bersifat non-invasif. Namun, kompleksitas sinyal EEG, keberadaan noise, artefak, serta variasi antarindividu masih menjadi tantangan utama dalam analisis dan klasifikasi ASD berbasis EEG. Penelitian ini bertujuan untuk mengkaji karakteristik neurofisiologis ASD melalui analisis parameter Hjorth pada domain waktu, mengembangkan prapemrosesan adaptif berbasis Butterworth–PID (BPID), serta menentukan teknik identifikasi dan klasifikasi ASD berbasis EEG melalui representasi fitur multiskala dan model temporal deep learning. Data EEG diperoleh secara primer dari 10 subjek, terdiri atas 5 individu dengan ASD dan 5 individu non-ASD/neurotipikal, menggunakan Cyton Biosensing Board. Analisis dilakukan melalui tiga tahap utama. Tahap pertama menganalisis karakteristik sinyal EEG menggunakan parameter Hjorth, yaitu Activity, Mobility, dan Complexity. Tahap kedua mengembangkan filter adaptif Butterworth–PID yang dioptimalkan menggunakan Bayesian Optimization dan diintegrasikan dengan arsitektur EEG-Inception serta State Space Models (SSM). Tahap ketiga mengevaluasi representasi fitur Principal Component Analysis (PCA) dan Multiscale Principal Component Analysis (MS-PCA) pada model RNN–BiLSTM dan TCN–Self Attention menggunakan skema subject-wise 5-fold cross-validation untuk mengurangi risiko subject leakage. Hasil penelitian menunjukkan bahwa parameter Hjorth Activity secara konsisten lebih rendah pada kelompok ASD dibandingkan kelompok non-ASD/neurotipikal, sedangkan Mobility dan Complexity relatif lebih serupa antar kelompok. Pada tahap klasifikasi berbasis temporal deep learning, kombinasi TCN–Self Attention dengan fitur MS-PCA mencapai rerata akurasi pengujian sebesar 97,97%. Selain itu, penerapan BPID mampu mereduksi kesalahan sinyal EEG sebesar 88%–95% dibandingkan filter Butterworth konvensional dan meningkatkan performa klasifikasi, dengan akurasi tertinggi sebesar 99,42% pada kombinasi BPID dan EEG-Inception. Temuan ini menunjukkan bahwa integrasi analisis karakteristik neurofisiologis, prapemrosesan adaptif, representasi fitur multiskala, dan temporal deep learning mampu menghasilkan kerangka identifikasi dan klasifikasi ASD berbasis EEG yang akurat, stabil, dan berpotensi mendukung pengembangan sistem bantu skrining awal yang objektif di masa mendatang.

The increasing prevalence of Autism Spectrum Disorder (ASD) in recent decades highlights the need for a more objective, measurable, and data-driven approach to early screening. Electroencephalography (EEG) is a promising neurophysiological modality because it records brain activity dynamics with high temporal resolution and is non-invasive. However, EEG signal complexity, noise, artifacts, and inter-individual variability remain major challenges in EEG-based ASD analysis and classification. This study aims to investigate the neurophysiological characteristics of ASD using time-domain Hjorth parameters, develop an adaptive Butterworth–PID (BPID) preprocessing method, and determine an effective EEG-based ASD identification and classification technique through multiscale feature representation and temporal deep learning models. Primary EEG data were acquired from 10 subjects, consisting of 5 individuals with ASD and 5 non-ASD/neurotypical individuals, using the Cyton Biosensing Board. The analysis was conducted in three main stages. The first stage analyzed EEG signal characteristics using Hjorth parameters, namely Activity, Mobility, and Complexity. The second stage developed an adaptive Butterworth–PID filter optimized using Bayesian Optimization and integrated with EEG-Inception and State Space Models (SSM). The third stage evaluated Principal Component Analysis (PCA) and Multiscale Principal Component Analysis (MS-PCA) feature representations using RNN–BiLSTM and TCN–Self Attention models under a subject-wise 5-fold cross-validation scheme to reduce the risk of subject leakage. The results showed that Hjorth Activity was consistently lower in the ASD group than in the non-ASD/neurotypical group, while Mobility and Complexity were relatively similar between groups. In the temporal deep learning-based classification stage, the combination of TCN–Self Attention with MS-PCA features achieved an average testing accuracy of 97.97%. Furthermore, the BPID approach reduced EEG signal error by 88%–95% compared with the conventional Butterworth filter and improved classification performance, with the highest accuracy of 99.42% achieved by the BPID and EEG-Inception combination. These findings indicate that integrating neurophysiological signal characterization, adaptive preprocessing, multiscale feature representation, and temporal deep learning can produce an accurate and stable EEG-based ASD identification and classification framework with potential application in objective early screening support systems.