Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions

Ziwei Wang; Zhentao He; Xingyi He; Hongbin Wang; Tianwang Jia; Jingwei Luo; Siyang Li; Xiaoqing Chen

cs.LGMar 11, 2026

Synthetic Data Generation for Brain-Computer Interfaces: Overview, Benchmarking, and Future Directions

Authors: Ziwei Wang, Zhentao He, Xingyi He, Hongbin Wang, Tianwang Jia, Jingwei Luo, Siyang Li, Xiaoqing Chen,

Organizations: School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China · Zhongguancun Academy, Beijing, China

Abstract

Deep learning has achieved transformative performance across diverse domains, largely driven by large-scale and high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a promising strategy to mitigate data scarcity, improve model generalization, and support data-efficient BCIs. This survey provides a comprehensive review of synthetic brain data generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, key applications, and future directions. We systematically categorize existing generation approaches into four types: signal-transformation-based, feature-based, model-based, and translation-based generation, and discuss their characteristics, advantages, and limitations. Furthermore, we benchmark representative brain signal generation approaches across four BCI paradigms, including motor imagery, epileptic seizure detection, steady-state visually evoked potentials, and auditory attention detection, to provide an objective comparison of their downstream utility. We also summarize evaluation principles for generated brain signals from multiple perspectives, including signal realism, physiological plausibility, downstream utility, and privacy preservation. Finally, we discuss the potential and challenges of current generation approaches and outline future research directions toward accurate, data-efficient, generalizable, and privacy-aware BCI systems. The benchmark codebase is available at https://github.com/wzwvv/DG4BCI.

Figures & tables

Figure 1: Data scarcity issue in BCIs, owing to the small data size, low signal quality, and/or significant differences across sessions, subjects, and devices.

Electrode Placement	Signal Type	Temporal Resolution	Spatial Resolution	Cost	Universality	Setup Duration
Non-invasive	EEG	High	Low	Cheap	High	Quick (Dry) / Moderate (Wet)
	fNIRS	Moderate	Moderate	Moderate	Moderate	Quick
	MEG	High	Moderate	Expensive	Low	Moderate
Partially Invasive	ECoG	High	High	Expensive	Low	Short-term implant
Invasive	SEEG	High	High	Expensive	Low	Short-term implant

Table 1: Characteristics of brain signals in BCIs.

Figure 2: Overall organization of this survey.

Figure 3: Data generation driven machine learning pipeline for BCIs, which includes brain signal acquisition, data preprocessing, data generation, feature engineering, and classification/regression. The latter two components can be unified into a single end-to-end neural network. Data generation approaches are categorized into four types: (a) signal-transformation-based generation, (b) feature-based generation, (c) model-based generation, and (d) translation-based generation.

Figure 4: Four types of data generation approaches for brain signals.

Type	Approach	Description	Parameter
Time domain	Noise Wang et al. (2018)	Adding uniform noise to an EEG trial in the time domain	$C_{noise}$ : the noise injection rate
	Mask Mohsenvand et al. (2020)	Masking a portion of EEG trials randomly	$C_{mask}$ : the masking rate
	Scale Zhang et al. (2022)	Scaling the voltage of EEG trials with a minor coefficient	$C_{scale}$ : the scaling rate
	Flip Zhang et al. (2022)	Performing voltage inversion	/
Frequency domain	FShift Zhang et al. (2022)	Shifting the frequency of EEG trials	$C_{fshift}$ : the frequency scaling rate
	FSurr Schwabedal et al. (2018)	Replacing the Fourier phases of trials with random numbers	$C_{p}$ : probability; $C_{m}$ : noise magnitude

Table 2: Four types of signal-transformation-based generation approaches.

Figure 5: Visualizations of brain signals before (blue lines) and after (red lines) eleven signal-transformation-based generation approaches, using 1-channel as an example.

Figure 6: Model-based generation approaches for brain signals, including GANS, VAEs, ARMs, and DDPMs.

Model Type	Approach	Signal Type	Paradigm	Dataset	Setup	Improvement
GAN	cc-LSTM-GAN Wen et al. (2023)	Spike	Motor execution	Two Monkeys	Conditional	16.0% accuracy
	DCGAN Fahimi et al. (2020)	EEG and ECG	Motor execution	Fourteen humans	Conditional	7.3% accuracy
	GDAL Ko et al. (2022)	EEG	MI	III-3a, III-4a, and IV-2a	Unconditional	2.0% accuracy
	SSVEP-GAN Figueiredo et al. (2023)	EEG	SSVEP	32 humans	Conditional	96s time-saving
	TOED-GAN Zeng et al. (2024)	EEG	SSVEP	Dial	Conditional	18.5% accuracy
	WGAN Aznan et al. (2021)	EEG	SSVEP	NAO	Unconditional	3.0% accuracy

Table 3: Comparison of existing studies on brain signal generative models, including GANs, VAEs, ARMs, and DDPMs.

Figure 7: Two types of translation-based generation for brain signals, taking two modalities of image and brain signal as an example.

Type	Strength	Limitation	Key Challenge
Signal-Transformation-Based	Uses domain-specific knowledge	Limited flexibility	Incorporates more brain signal features
Signal-Transformation-Based	High interpretability	Struggles with nonlinear data	Ensures diversity of generated signals
Feature-Based	Generates diverse samples	Lack of direct signal fidelity	Maintains feature integrity
Feature-Based	Useful for imbalanced data	Overfits to the minority	Enhances model robustness
Model-Based	Handles high-dimensional data	Computationally expensive	Manages model complexity
Model-Based	Flexible and adaptable	Prone to mode collapse	Efficient training

Table 4: Comparison of four types of synthetic data generation approaches in BCIs.

Paradigm	Dataset	Number of	Number of	Sampling	Trial Length	Number of	Task Types
Paradigm	Dataset	Subjects	EEG Channels	Rate (Hz)	(seconds)	Trials	Task Types
MI	IV-2a	9	22	250	4	1,296	left/right hand
	Zhou2016	4	14	250	5	409	left/right hand
	Blankertz2007	7	59	250	3	1,400	left/right hand or left hand/right foot
	BNCI2014002	14	15	512	5	1,400	right hand/both feet
	BNCI2015001	12	13	512	5	2,400	right hand/both feet

Table 5: Summary of the five MI datasets, two seizure datasets, two SSVEP datasets, and two AAD datasets. The sampling rate and trial length refer to the preprocessed signals used in our experiments.

Backbone	Dataset	None	Flip	Noise	Scale	FShift	FSurr	HS	CR	DWTaug
SCNN	IV-2a	72.22 ±1.03	75.10 ±0.95	74.44 ±0.74	74.39 ±0.51	74.95 ±0.95	65.92 ±1.24	73.37 ±1.05	74.76 ±1.10	75.56 ±0.93
	Zhou2016	81.97 ±1.63	81.82 ±0.74	82.62 ±1.20	83.67 ±0.91	81.67 ±0.61	63.04 ±1.20	81.57 ±0.58	81.87 ±0.78	84.92 ±1.32
	Blankertz2007	70.04 ±0.79	74.10 ±0.90	73.59 ±0.90	71.96 ±0.87	73.57 ±0.44	61.51 ±0.88	73.38 ±0.63	73.89 ±0.66	74.69 ±1.17
	BNCI2014002	68.13 ±0.74	72.10 ±0.37	72.97 ±0.41	72.36 ±0.74	72.28 ±0.70	66.74 ±1.14	71.72 ±0.67	72.43 ±0.45	72.21 ±1.01
	BNCI2015001	69.71 ±0.67	69.80 ±0.89	69.31 ±1.00	68.98 ±0.82	69.35 ±0.90	68.60 ±0.65	68.32 ±0.96	69.50 ±0.73	74.11 ±0.29
	Average	72.41	74.58	74.59	74.27	74.36	65.16	73.67	74.49	76.30

Table 6: Average classification accuracies (%) on five MI datasets with five decoding models under the cross-dataset setting. The best average performance of each network is marked in bold, and the second best by an underline.

Dataset	Backbone	None	Flip	Noise	Scale	FShift	HS	CR
CHSZ	EEGNet	81.92 ±0.76	79.46 ±0.78	80.63 ±0.81	81.74 ±0.31	80.43 ±0.93	80.40 ±1.19	82.77 ±0.40
	EEGWaveNet	78.05 ±0.34	79.03 ±0.70	78.21 ±0.97	79.02 ±0.02	77.65 ±1.31	75.25 ±1.09	79.47 ±0.57
	Average	79.99	79.25	79.42	80.38	79.04	77.83	81.12
NICU	EEGNet	61.22 ±0.54	60.43 ±0.24	58.88 ±0.29	58.69 ±0.43	58.79 ±0.51	57.17 ±0.46	62.03 ±0.40
	EEGWaveNet	62.86 ±0.68	61.17 ±0.47	62.88 ±0.50	65.12 ±0.71	63.38 ±0.53	63.39 ±0.23	66.04 ±0.16
	Average	62.04	60.80	60.88	61.91	61.09	60.28	64.04

Table 7: Average classification BCAs (%) on two seizure datasets with EEGNet and EEGWaveNet models under the cross-subject setting. The best average performance of each network is marked in bold, and the second best by an underline.

Dataset	Backbone	None	Flip	Noise	Scale	FShift	DWTaug
Nakanishi	FBCCA	68.06 ±0.68	7.90 ±0.57	67.99 ±0.69	68.06 ±0.68	60.68 ±2.21	72.86 ±3.37
	EEGNet	45.69 ±3.71	35.90 ±1.13	49.88 ±2.79	49.14 ±2.82	48.63 ±3.94	51.93 ±2.70
	CCNN	70.54 ±1.74	69.14 ±3.49	70.65 ±3.33	70.81 ±3.41	70.08 ±2.27	71.36 ±3.10
	SSVEPNet	90.03 ±1.20	82.89 ±0.87	90.18 ±1.49	90.22 ±1.39	88.65 ±1.38	90.88 ±1.08
	Average	68.58 ±1.83	48.96 ±1.52	69.68 ±2.08	69.56 ±2.08	67.01 ±2.45	71.76 ±2.56
Benchmark	FBCCA	63.08 ±1.06	2.38 ±0.42	62.83 ±1.53	63.08 ±1.06	54.20 ±3.07	71.42 ±0.66

Table 8: Average classification accuracies (%) on two SSVEP datasets with four decoding backbones under the within-subject setting. The best average performance of each network is marked in bold, and the second best by an underline.

Dataset	Backbone	None	Flip	Noise	Scale	FShift	FSurr	CR	DWTaug
KUL	IFNet	90.46 ±0.25	84.76 ±2.13	90.73 ±0.36	90.92 ±0.27	92.77 ±0.49	91.43 ±0.30	91.17 ±0.46	91.39 ±0.19
	DBConformer	84.56 ±0.36	84.26 ±1.12	85.20 ±0.66	85.38 ±0.94	88.56 ±0.34	85.05 ±0.36	85.63 ±0.43	86.35 ±0.66
	DARNet	89.60 ±0.72	93.58 ±0.30	89.81 ±0.43	89.85 ±0.30	90.80 ±0.45	89.89 ±0.38	89.67 ±0.59	90.45 ±0.45
	Average	88.21	87.53	88.58	88.72	90.71	88.79	88.82	89.40
DTU	IFNet	82.25 ±0.75	76.77 ±0.63	81.98 ±0.54	82.48 ±0.48	83.09 ±0.45	83.32 ±0.39	80.30 ±0.32	82.50 ±0.34
	DBConformer	80.42 ±0.32	79.93 ±0.81	80.74 ±0.68	80.98 ±0.55	80.87 ±0.29	80.98 ±0.20	79.47 ±0.86	81.38 ±0.60

Table 9: Average classification accuracies (%) on two AAD datasets with IFNet, DBConformer, and DARNet models under the within-subject setting. The best average performance of each network is marked in bold, and the second best by an underline.

Dataset	Backbone	None	Vanilla			GAN			VAE
Dataset	Backbone	None	CNN	CNN-Trans.	CNN-LSTM	CNN	CNN-Trans.	CNN-LSTM	CNN	CNN-Trans.	CNN-LSTM
Nakanishi	FBCCA	79.50 ±4.0	79.44 ±4.01	79.50 ±4.2	79.78 ±4.3	86.06 ±4.2	86.33 ±3.0	85.94 ±3.4	79.50 ±4.1	79.61 ±4.2	79.72 ±4.4
	EEGNet	90.50 ±3.9	87.56 ±2.4	90.67 ±2.4	88.00 ±2.2	89.22 ±4.0	89.83 ±1.7	90.39 ±2.1	88.17 ±2.9	90.61 ±2.0	89.61 ±1.2
	CCNN	91.33 ±2.6	90.00 ±4.5	91.17 ±2.6	89.78 ±2.7	92.50 ±3.4	92.83 ±2.6	92.83 ±2.6	90.00 ±4.5	91.11 ±2.8	89.83 ±2.6
	SSVEPNet	97.17 ±0.7	97.33 ±1.3	96.61 ±1.3	96.89 ±0.6	97.06 ±1.4	97.22 ±1.5	97.06 ±1.0	97.22 ±1.4	97.22 ±1.0	97.17 ±1.3
	Average	89.63	88.58	89.49	88.61	91.21	91.55	91.56	88.72	89.64	89.08

Table 10: Average classification accuracies (%) on two SSVEP datasets with four base classifiers under the within-subject setting, comparing nine generative models. The best average performance of each network is marked in bold, and the second best by an underline.

Figure 8: Evaluation framework for generated data in BCIs, encompassing data reliability, quality, task performance, model performance, multimodal consistency, and privacy preservation ability.

Synthetic Data Generation

Synthetic data generation creates artificial examples, simulations, and datasets to train, test, or protect machine-learning systems.

See all 606 papers

A Survey on Bridging EEG Signals and Generative AI: From Image and Text to Beyond
Feb 17, 2025Shreya Shukla, Jose Torres, Akshaj Murhekar +4EEG Decoding EEG-Based Visual Decoding
School of Information, The University of Texas at Austin, Austin, TX, USA · Institute of Applied Computer Science, Łód´z University of Technology, Łód´z, Poland
Riemannian Geometry-Preserving Variational Autoencoder for MI-BCI Data Augmentation
Mar 11, 2026Viktorija Poļaka, Ivo Pascal de Jong, Andreea Ioana SburleaSynthetic Data Augmentation Variational Autoencoders
Faculty of Science and Engineering, University of Groningen, Groningen, The Netherlands
UniBCI: Towards a Unified Pretrained Model for Invasive Brain-Computer Interfaces
Apr 30, 2026Binjie Hong, Rui Xiong, Liyuan Han +1Brain-Computer Interfaces Self-Supervised Pre-Training
Brain Science Data Center · CEBSIT, Institute of Neuroscience

Feb 17, 2025J/K move · Enter open · S save

A Survey on Bridging EEG Signals and Generative AI: From Image and Text to Beyond

Shreya Shukla, Jose Torres, Akshaj Murhekar, Christina Liu+3

School of Information, The University of Texas at Austin, Austin, TX, USA · Institute of Applied Computer Science, Łód´z University of Technology, Łód´z, Poland

EEG Decoding EEG-Based Visual Decoding Conditional Generative Modeling

Open paper AskPDF arXiv

Decoding neural activity into human-interpretable representations is a key research direction in brain-computer interfaces (BCIs) and computational neuroscience. Recent progress in machine learning and generative AI has driven growing interest in transforming non-invasive Electroencephalography (EEG) signals into images, text, and audio. This survey consolidates and analyzes developments across EEG-to-image synthesis, EEG-to-text generation, and EEG-to-audio reconstruction. We conducted a structured literature search across major databases (2017-2025), extracting key information on datasets, generative architectures (GANs, VAEs, transformers, diffusion models), EEG feature-encoding techniques, evaluation metrics, and the major challenges shaping current work in this area. Our review finds that EEG-to-image models predominantly employ encoder-decoder architectures built on GANs, VAEs, or diffusion models; EEG-to-text approaches increasingly leverage transformer-based language models for open-vocabulary decoding; and EEG-to-audio methods commonly map EEG signals to mel-spectrograms that are subsequently rendered into audio using neural vocoders. Despite promising advances, the field remains constrained by small and heterogeneous datasets, limited cross-subject generalization, and the absence of standardized benchmarks. By consolidating methodological trends and available datasets, this survey provides a foundational reference for advancing EEG-based generative AI and supporting reproducible research. We further highlight open-source datasets and baseline implementations to facilitate systematic benchmarking and accelerate progress in EEG-driven neural decoding.

Abstract

Synthetic Data Generation

Explore similar work

A Survey on Bridging EEG Signals and Generative AI: From Image and Text to Beyond

Riemannian Geometry-Preserving Variational Autoencoder for MI-BCI Data Augmentation

UniBCI: Towards a Unified Pretrained Model for Invasive Brain-Computer Interfaces