GEPD: GAN-Enhanced Generalizable Model for EEG-Based Detection of Parkinson’s Disease

Qian Zhang, Ruilin Zhang, Biaokai Zhu, Jun Xiao, Yifan Liu, Xun Han, Zhe WangLecture Notes in Computer Science 2025 · DOI · OpenAlex

Abstract

Electroencephalography (EEG) has been established as an effective method for detecting Parkinson's disease (PD). However, the variability across different EEG datasets and the small size of available data pose significant challenges for training models that generalize to cross-dataset scenarios. This paper proposes GEPD, a GAN-enhanced generalizable model specifically for EEG-based cross-dataset classification of PD. GEPD utilizes a generative network to create fusion EEG data by controlling distribution similarity and an EEG signal quality assessment model to ensure high-quality data generation. A classification network then captures time-frequency characteristics using multiple convolutional layers. Evaluation results show that GEPD achieves an accuracy of 84.3% and an F1-score of 84.0% in cross-dataset settings, demonstrating its strong generalizability.

Introduction

The early detection of Parkinson's Disease (PD) is critical for effective management, and Electroencephalography (EEG) offers a non-invasive, cost-effective diagnostic tool. Despite the success of deep learning in this field, two primary obstacles remain: data scarcity and poor generalization. Medical privacy and high collection costs result in small datasets, leading to overfitting. Furthermore, differences in equipment and subject conditions across clinical sites create a "domain shift," where models trained on one dataset fail on another. GEPD addresses these issues by combining generative data augmentation with a robust, generalizable classification architecture.

Methodology: The GEPD Framework

The GEPD pipeline consists of three main stages: data generation, channel pruning, and classification.

Overview of the GEPD framework, illustrating the generation and classification stages.

1. Data Augmentation (PD-WGAN)

To mitigate the small-data problem, GEPD employs a Wasserstein GAN (WGAN) enhanced with Convolutional Block Attention Modules (CBAM). This generative network, PD-WGAN, creates synthetic healthy and PD EEG signals from random noise. These synthetic signals are combined with real data to form a "fusion" dataset, providing a richer training set for the classifier.

2. Quality Assessment and Channel Pruning

Generating realistic time-series data like EEG is challenging. GEPD uses an LSTM-based autoencoder to assess the quality of generated signals via reconstruction error. To further refine the input, the model implements a channel-pruning technique based on Jensen-Shannon (JS) divergence.

The channel pruning process based on Jensen-Shannon divergence to filter informative and high-fidelity EEG channels.

As shown in Figure 2, the pruning process evaluates two criteria:

  • Informativeness: Channels must show significant divergence between healthy and PD groups to be useful for classification.
  • Fidelity: Channels must show low divergence between real and generated data to ensure the synthetic data is reliable.

By pruning the input from 60 to 38 channels, the model focuses on the most discriminative and high-fidelity brain regions, reducing noise and improving generalization.

3. Classification (PDNex)

The final stage uses PDNex, a lightweight network designed for efficiency and generalization. It utilizes depthwise separable and dilated convolutions to extract spatial and temporal features. Regularization techniques, such as cosine annealing, are employed to prevent the model from overfitting to the source dataset.

Experimental Results

The model was trained on the UI Dataset (28 participants) and tested on the entirely unseen UNM Dataset (54 participants) to evaluate zero-shot cross-dataset performance.

Key Performance Metrics

GEPD achieved state-of-the-art results in the cross-dataset setting:

  • Accuracy: 84.3%
  • F1-Score: 84.0%

These results significantly outperform baseline models like EEGNex (75.3% accuracy) and EEG-ITNet (73.4% accuracy). Ablation studies confirmed the importance of each component: removing GAN-generated data led to a 2.9% drop in accuracy, while removing channel pruning resulted in a 2.0% decrease. Additionally, the researchers identified an optimal augmentation scale, finding that performance peaked when the original dataset was expanded by 1.7 times with synthetic data.

Conclusion

GEPD demonstrates that "less is more" in EEG analysis; by strategically pruning channels and augmenting data with a quality-controlled GAN, the model achieves superior generalization. Future work will explore more advanced generative architectures, such as diffusion models, and focus on enhancing model interpretability to better understand the pathological markers of Parkinson's Disease.

Summary updated

BibTeX
@inproceedings{zhang2025gepd,
  author = {Qian Zhang and Ruilin Zhang and Biaokai Zhu and Jun Xiao and Yifan Liu and Xun Han and Zhe Wang},
  title = {{GEPD: GAN-Enhanced Generalizable Model for EEG-Based Detection of Parkinson’s Disease}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2025},
  doi = {10.1007/978-981-95-0033-8\_26},
}

Loading related content...

Figure