Synthetic SSVEP Dataset Generated with GAN and Diffusion Models

This repository provides synthetic steady-state visual evoked potential (SSVEP) EEG signals generated using two generative approaches: Generative Adversarial Networks (GAN) and Diffusion Models (DM). The dataset accompanies the article:

Przemysław Wiszniewski, Marcin Kołodziej, Andrzej Majkowski, Comparison of Generative Adversarial Networks and Diffusion Models for Synthetic Steady-State Visual Evoked Potential Signal Generation, DOI: to be added

The dataset contains both real training EEG segments and synthetic signals generated by trained generative models.

Dataset Overview

The dataset includes synthetic SSVEP signals generated for:

5 subjects

4 stimulation frequencies: 5 Hz, 6 Hz, 7 Hz, 8 Hz

2 generative models:

GAN

Diffusion Model (DM)

Each signal represents a 1-second EEG segment sampled at 256 Hz, therefore each sample contains 256 time samples.

Synthetic signals were generated independently for each subject and each stimulation frequency.

Data Acquisition

The original EEG data were recorded using:

amplifier: g.USBAmp 2.0 (g.tec)

resolution: 24-bit A/D converter

sampling frequency: 256 Hz

electrodes: 16 electrodes according to the international 10-20 system

Visual stimulation was provided using a green LED placed approximately 1 meter from the participant and driven by a function generator at the following frequencies: 5 Hz, 6 Hz, 7 Hz, 8 Hz.

For the analysis presented in the article, the Oz electrode was used due to its strong SSVEP response.

Participants included five subjects aged 23–46.

Preprocessing

The EEG signals were preprocessed using the following steps:

Butterworth bandpass filter: 0.1–100 Hz

notch filter: 48–51 Hz to suppress power-line interference

normalization to the range [-1, 1] before training generative models

Generative Models

Two generative approaches were used to produce synthetic EEG signals:

Generative Adversarial Network (GAN)

A fully connected BasicGAN architecture consisting of:

Generatorlatent dimension: 20

Discriminatortwo hidden layers with ReLU activation

The training objective combined time-domain and spectral losses, with stronger emphasis on spectral fidelity.

Diffusion Model (DM)

A 1D Denoising Diffusion Probabilistic Model (DDPM) was used with a modified U-Net architecture designed for temporal signals.

The model progressively adds Gaussian noise to the data during the forward process and learns to reconstruct signals during the reverse denoising process.

Repository Structure

The dataset is organized into two main directories corresponding to the generative models.

DM/GAN/DM.zipGAN.zip

Each directory contains files corresponding to a specific subject and stimulation frequency.

Example structure:

DM/sub1_5Hz.matsub1_6Hz.mat...sub5_8Hz.matGAN/sub1_5Hz.matsub1_6Hz.mat...sub5_8Hz.mat

File naming convention:

subX_YHz.mat

where:

X – subject index (1–5)

Y – stimulation frequency (5, 6, 7, or 8 Hz)

File Format

Each .mat file contains two variables:

generated 1000 × 256 doubleoriginal 348 × 256 double

Meaning:

original

Real EEG segments used to train the generative models.

348 samples256 time points (1 second at 256 Hz)

generated

Synthetic EEG segments produced by the trained generative model.

1000 samples256 time points (1 second at 256 Hz)

Rows correspond to individual EEG epochs, while columns represent time samples.

Example MATLAB Usage

load('sub1_6Hz.mat')orig = original; % [348 x 256]gen = generated; % [1000 x 256]fs = 256;t = (0:255)/fs;plot(t, orig(1,:))xlabel('Time [s]')ylabel('Amplitude')

Intended Use

The dataset can be used for:

evaluation of generative models for EEG signals

benchmarking synthetic SSVEP data generation

signal processing and spectral analysis

development and testing of SSVEP-BCI algorithms

machine learning and classification experiments