Electronic Theses and Dissertation

Penyakit Parkinson (PD) merupakan gangguan neurodegeneratif progresif akibat kerusakan sel saraf dopaminergik di substantia nigra, yang menyebabkan penurunan kadar dopamin serta gangguan motorik dan non-motorik seperti tremor, kekakuan otot, dan gangguan tidur. Deteksi dini Parkinson masih menjadi tantangan karena gejala awalnya tidak spesifik dan metode diagnostik klinis bersifat subjektif. Penelitian ini mengusulkan model klasifikasi penyakit Parkinson berbasis sinyal EEG dengan mengintegrasikan Independent Component Analysis (ICA) untuk pembersihan sinyal dan Convolutional Neural Network satu dimensi (CNN Conv1D) untuk ekstraksi fitur serta klasifikasi. Eksperimen dilakukan pada dua skenario: dataset tunggal (UC San Diego, Universitas Iowa, dan Universitas New Mexico) dan dataset gabungan yang mengombinasikan ketiganya. Hasil menunjukkan bahwa variasi sumber data berpengaruh terhadap kinerja dan generalisasi model. Dataset tunggal menghasilkan akurasi pelatihan tinggi namun cenderung overfitting, sementara dataset gabungan memberikan hasil lebih stabil dengan akurasi pelatihan 97,61% dan validasi 87,83%. Selain itu, hasil evaluasi menunjukan dataset gabungan memperoleh akurasi 91.31%, presisi 91.54%, recall 95.21%, dan F1-Score 93.35%. Penggunaan optimizer RMSprop meningkatkan akurasi validasi hingga 90,83% dengan loss yang lebih rendah. Selain itu, dataset gabungan dengan optimizer RMSProp juga memperoleh akurasi 92.20%, presisi 95.18%, recall 92.56%, dan F1-Score 93.86%. Secara keseluruhan, pendekatan ICA–CNN Conv1D pada data EEG multi-sumber terbukti meningkatkan stabilitas dan keandalan model, serta berpotensi mendukung deteksi dini Parkinson secara lebih objektif dan akurat.

Parkinson’s disease (PD) is a progressive neurodegenerative disorder caused by damage to dopaminergic neurons in the substantia nigra, resulting in decreased dopamine levels and motor and non-motor impairments such as tremor, muscle rigidity, and sleep disturbances. Early detection of PD remains challenging because initial symptoms are nonspecific and clinical diagnostic methods are subjective. This study proposes an EEG-based Parkinson’s disease classification model by integrating Independent Component Analysis (ICA) for signal denoising and a one-dimensional Convolutional Neural Network (Conv1D CNN) for feature extraction and classification. Experiments were conducted under two scenarios: single-source datasets (UC San Diego, University of Iowa, and University of New Mexico) and multi-source dataset integrating all three sources. The results demonstrate that data source variability significantly affects model performance and generalization. Models trained on single-source datasets achieved high training accuracy but tended to overfit, whereas the multi-source dataset produced more stable performance with a training accuracy of 97.61% and a validation accuracy of 87.83%. Furthermore, evaluation results show that the multi-source dataset achieved an accuracy of 91.31%, precision of 91.54%, recall of 95.21%, and an F1-score of 93.35%. The use of the RMSprop optimizer further improved validation accuracy to 90.83% with lower loss. Additionally, the multi-source dataset with the RMSprop optimizer achieved an accuracy of 92.20%, precision of 95.18%, recall of 92.56%, and an F1-score of 93.86%. Overall, the ICA–Conv1D CNN approach using multi-source EEG data enhances model stability and reliability, and shows strong potential to support more objective and accurate early detection of Parkinson’s disease.