Electronic Theses and Dissertation

Electroencephalography (EEG) banyak digunakan karena bersifat non-invasif dan memiliki resolusi temporal tinggi. Namun, sinyal EEG mentah sangat rentan terhadap artefak biologis dan noise lingkungan, sehingga diperlukan proses denoising yang efektif sebelum klasifikasi. Penelitian ini bertujuan mengevaluasi pengaruh dua metode denoising multiskala, yaitu Multiscale Principal Component Analysis (MS-PCA) dan Multiscale Independent Component Analysis (MS-ICA), terhadap kinerja klasifikasi sinyal EEG penyandang ASD. Dataset EEG diperoleh melalui akuisisi 16 kanal dari 16 subjek, yang terdiri atas 8 subjek ASD dan 8 subjek normal, menggunakan OpenBCI Cyton Board. Tahap pre-processing dilakukan menggunakan Butterworth Band-Pass Filter pada rentang 0,5-40 Hz, dilanjutkan dengan penerapan MS-PCA dan MS-ICA untuk mereduksi artefak. Sinyal hasil pre-processing kemudian disegmentasi menjadi epoch berdurasi 2 detik dengan overlap 50%. Proses klasifikasi dilakukan menggunakan dua pendekatan, yaitu Light Gradient Boosting Machine (LightGBM) sebagai representasi machine learning dan Temporal Convolutional Network with Self-Attention (TCN-SA) sebagai representasi deep learning. Pada LightGBM dilakukan ekstraksi fitur Power Spectral Density (PSD) dan Band Power, sedangkan pada TCN-SA segmen EEG hasil pre-processing digunakan langsung tanpa ekstraksi fitur manual. Evaluasi dilakukan menggunakan 8-fold subject-wise cross-validation dengan metrik akurasi, presisi, recall, F1-score, spesifisitas, dan Area Under the Curve (AUC). Hasil penelitian menunjukkan bahwa kombinasi LightGBM + MS-PCA memberikan kinerja terbaik dengan rata-rata akurasi 99,14% dan AUC 99,96%. Sementara itu, LightGBM + MS-ICA menghasilkan akurasi 98,85%, TCN-SA + MS-PCA 95,03%, dan TCN-SA + MS-ICA 94,70%. Analisis statistik berpasangan pada level subjek menunjukkan bahwa MS-PCA cenderung lebih baik daripada MS-ICA pada kedua model. Hasil ini menunjukkan bahwa pada dataset dan protokol evaluasi yang digunakan, MS-PCA lebih efektif dalam mendukung klasifikasi EEG penyandang ASD.

Electroencephalography (EEG) is widely used due to its non-invasive nature and high temporal resolution. However, raw EEG signals are highly susceptible to biological artifacts and environmental noise, requiring an effective denoising process before classification. This study aims to evaluate the effect of two multiscale denoising methods, namely Multiscale Principal Component Analysis (MS-PCA) and Multiscale Independent Component Analysis (MS-ICA), on the classification performance of EEG signals from individuals with ASD. The EEG dataset was obtained through 16-channel acquisition from 16 subjects, consisting of 8 ASD subjects and 8 normal subjects, using the OpenBCI Cyton Board. The pre-processing stage was carried out using a Butterworth Band-Pass Filter in the range of 0.5-40 Hz, followed by the application of MS-PCA and MS-ICA to reduce artifacts. The pre-processed signal was then segmented into 2-second epochs with 50% overlap. The classification process was carried out using two approaches, namely the Light Gradient Boosting Machine (LightGBM) as a representation of machine learning and Temporal Convolutional Network with Self-Attention (TCN-SA) as a representation of deep learning. In LightGBM, Power Spectral Density (PSD) and Band Power features were extracted, while in TCN-SA, the pre-processed EEG segments were used directly without manual feature extraction. Evaluation was carried out using 8-fold subject-wise cross-validation with metrics of accuracy, precision, recall, F1-score, specificity, and Area Under the Curve (AUC). The results showed that the combination of LightGBM + MS-PCA provided the best performance with an average accuracy of 99.14% and an AUC of 99.96%. Meanwhile, LightGBM + MS-ICA produced an accuracy of 98.85%, TCN-SA + MS-PCA 95.03%, and TCN-SA + MS-ICA 94.70%. Pairwise statistical analysis at the subject level showed that MS-PCA tended to outperform MS-ICA in both models. These results indicate that, given the dataset and evaluation protocol used, MS-PCA is more effective in supporting EEG classification of individuals with ASD.