سامانه همایش‌ها TSTA | ثبت‌نام و ارسال مقاله کنفرانس‌ها و کنگره‌های ملی

صفحه اصلی / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Parkinson’s Disease Classification Using EEG and a Hybrid EEGNet–LSTM Architecture

نویسندگان :

Pouya Taghipour Langrodi¹ Amirsadra Khodadadi² Ali Sadat Modaresi³ Mohammad Ahadzadeh⁴ Mostafa Rostami⁵ Sadegh Madadi⁶

1- Professor, Department of Mechanical Engineering, Amirkabir University of Technology (Tehran Polytechnic) 2- Professor, Department of Mechanical Engineering, Amirkabir University of Technology (Tehran Polytechnic) 3- Professor, Department of Mechanical Engineering, Amirkabir University of Technology (Tehran Polytechnic) 4- Professor, Department of Mechanical Engineering, Amirkabir University of Technology (Tehran Polytechnic) 5- Professor, Department of Mechanical Engineering, Amirkabir University of Technology (Tehran Polytechnic) 6- Professor, Department of Mechanical Engineering, Amirkabir University of Technology (Tehran Polytechnic)

کلمات کلیدی :

Parkinson’s Disease،Electroencephalography،Machine Learning،Simon Conflict،Deep Neural Networks

چکیده :

Parkinson's disease (PD) is a common progressive neurodegenerative disorder that causes motor problems and cognitive-control problems that slowly get worse over time. These problems often show up years before a clinical diagnosis. To meet the need for objective early biomarkers, high-density electroencephalography (EEG) was recorded from 56 subjects (28 PD patients and 28 controls) while they did the Simon Conflict Task 200 times. This task tests how well people can stop themselves from responding when the conditions are the same or different. After a few preprocessing steps, which included 0.1–40 Hz band-pass filtering, common-average re-referencing, and independent component analysis (ICA) with ICLabel-guided artifact rejection, one-second epochs that were time-locked to the start of the stimulus were taken out. We then created a hybrid deep-learning framework that combined EEGNet for spatial feature extraction across 64 channels with three stacked bidirectional Long Short-Term Memory (LSTM) layers to capture temporal dynamics. Three shallow supervised models were used to classify the 64-dimensional spatiotemporal representations for each epoch: support vector machine (SVM), k-nearest neighbors (kNN), and an ensemble of SVM and Naïve Bayes. SVM did the best, with 89.7% accuracy, 91.8% sensitivity, and 85.0% specificity. This was a 5–10% improvement over traditional handcrafted-feature classifiers (p < 0.01). These results show that end-to-end spatial-temporal feature learning from task-evoked EEG is a powerful, non-invasive way to accurately separate Parkinson’s patients and the control group.