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

Feature-Conditioned WGAN for Generating Alzheimer’s EEG: Enabling GAN-Based Synthesis Under Data Scarcity

نویسندگان :

Parsa Bahramsari¹ Alireza Taheri²

1- Social and Cognitive Robotics Lab, Department of Mechanical Engineering, Sharif University of Technology, Tehran, Iran 2- Social and Cognitive Robotics Lab, Department of Mechanical Engineering, Sharif University of Technology, Tehran, Iran

کلمات کلیدی :

Alzheimer’s disease‎،Electroencephalography‎،Conditional Wasserstein GAN‎،Feature matching،Synthetic data generation

چکیده :

Alzheimer’s disease (AD) significantly impairs cognitive ‎function, making early detection and personalized care ‎crucial. Electroencephalography (EEG) provides a non-‎invasive, low-cost window into cortical oscillations and is ‎sensitive to AD-related spectral slowing and reduced ‎temporal complexity. However, acquiring high-quality EEG ‎data is often limited by factors such as patient fatigue, ‎session variability, and logistical challenges, especially in ‎environments like socially assistive robots (SARs). These ‎constraints make it difficult to gather sufficient data for ‎training reliable deep models for AD detection. To address ‎this challenge, we propose a feature-conditioned ‎Wasserstein generative adversarial network (fc-WGAN) ‎that generates class and subject specific EEG segments ‎from minimal training data. We first analyze a broad set of ‎time-domain and frequency-domain EEG features to ‎identify those most discriminative between AD and ‎cognitively normal groups. Notably, features like nonlinear ‎energy and band powers consistently demonstrate high ‎separability. fc-WGAN aligns the mean and variance of ‎these features between real and generated EEG batches, ‎enhancing physiological realism and class consistency. ‎Starting from only 200 overlapping 3-second segments per ‎subject, our method improves EEGNet classification ‎accuracy from 87.5±4.5% to 96.2±4.4% by effectively ‎augmenting the training dataset. These results underscore ‎the power of feature-aligned generation in overcoming data ‎scarcity and demonstrate the practical utility of fc-WGAN ‎for SAR-based cognitive assessment and early AD detection ‎in real-world settings. ‎