Generative diffusion meets domain adaptation: a framework for EEG cross-subject motor imagery classification

Abstract

Cross-subject motor imagery classification remains challenging due to EEG data scarcity and inter-subject variability. This study proposes a novel framework integrating generative data augmentation with domain adaptation. First, we employ a diffusion probabilistic model to generate high-fidelity synthetic EEG samples, effectively enriching the training data. Subsequently, we propose the AMSC-DANN architecture, which synergizes an Adaptive Multi-Scale Convolution (AMSC) module for extracting multi-granular features with a Domain Adversarial Neural Network (DANN). This combination enables the model to learn discriminative temporal-spectral representations while simultaneously aligning feature distributions across different subjects. Extensive experiments on BCI Competition IV datasets 2a and 2b demonstrate that our proposed framework outperforms state-of-the-art baselines, validating its effectiveness in enhancing cross-subject generalization.

Introduction

Brain-computer Interface (BCI), as a cutting-edge human-computer interaction technology, aims to establish a direct communication path between the brain and external devices [1]. In recent years, BCI technology has demonstrated great application potential in fields such as medical rehabilitation, transportation, device control, and entertainment, attracting extensive attention from both the academic and industrial communities [2]. Motor Imagery (MI) stands as one of the most prevalent paradigms in BCI. During the imagination of body movements, neural oscillations in the sensorimotor cortex undergo event-related desynchronization (ERD) or synchronization (ERS) [3]. These variations in neural activity can be effectively captured by Electroencephalography (EEG) devices [4].

EEG-based BCI systems primarily comprise three major types of stimulation paradigms: event-related potentials [5], steady-state visual evoked potentials [6], and sensorimotor rhythms [7]. Among them, Motor Imagery EEG (MI-EEG) is recognized as a vital paradigm for motor rehabilitation, as it effectively captures the sensorimotor rhythms induced by the imagination of motor tasks. The rise of deep learning (DL) has significantly improved the performance of Motor Imagery (MI) decoding. Unlike traditional methods relying on manual feature extraction like CSP [8], DL models such as CNNs [9], EEGNet [10], can automatically learn complex representations from raw EEG data. However, their performance often deteriorates significantly in cross-subject tasks. First, there is the common problem of the scarcity of EEG training data [11]. Since collecting large-scale, high-quality EEG datasets is hindered by time constraints and subject fatigue, most research suffer from a severe scarcity of labeled samples. This data limitation causes deep learning models to overfit and fail in learning features. Second, the inherent non-stationarity of EEG signals and the individual differences among subjects also affect the accuracy of recognition. Due to diverse physiological states, unique cortical folding, and varying electrode impedances, EEG patterns often exhibit drastic distribution shifts across different subjects or even different sessions of the same subject [12]. This phenomenon, technically referred to as domain shift, inevitably leads to performance deterioration when a model trained on existing subjects is applied to an unseen subject. Consequently, relying solely on data expansion is insufficient to guarantee universality.

The main contributions of this paper can be summarized as follows:

1. We propose an integrated framework that innovatively combines generative diffusion models with discriminative domain adaptation. This hybrid strategy simultaneously addresses the dual bottlenecks of data scarcity and individual variability in BCI systems.

2. We design an Adaptive Multi-Scale Convolution with Domain Adversarial Neural Network (AMSC-DANN). By combining adaptive multi-scale feature extraction with domain adversarial learning, this module effectively aligns the feature distribution among different subjects.

3. We conduct relevant experiments on the motor imagery datasets to demonstrate the effectiveness of this framework in the cross-subject classification.

The rest of the paper is organized as follow: Sect. 2 shows the related work. Section 3 introduces our proposed framework. Section 4 states the experimental datasets, experimental settings, and classification results. Section 5 shows three ablation studies of our proposed work. In section 6, we provide a discussion on the proposed framework, highlighting its distinct merits while acknowledging its current constraints. Section 7 concludes the paper and points out the future works.

Related work

Convolutional neural networks in motor imagery decoding

In recent years, Deep Learning (DL) has achieved remarkable success in Motor Imagery classification, gradually superseding traditional machine learning methods based on manual feature engineering, such as Filter Bank Common Spatial Pattern (FBCSP) [13]. Convolutional Neural Networks (CNNs) [14], in particular, have become the dominant architecture due to their powerful capability in automatically extracting spatial-temporal representations from raw EEG signals. Representative models, including EEGNet [10], ShallowConvNet and DeepConvNet [14], utilize compact convolutional kernels to capture local dependencies and frequency-specific patterns. [15] introduced a dynamic graph neural network model, MCGNet+, which improves motor imagery classification accuracy by leveraging mutual information to construct evolving brain connection graphs. However, a fundamental limitation of these standard architectures lies in their reliance on fixed receptive fields (kernel sizes). EEG signals are inherently non-stationary, characterized by complex dynamics that span multiple frequency bands and varying temporal scales. Consequently, employing a single, static kernel size restricts the model’s ability to simultaneously capture broad contextual information and fine-grained local transient features.

Emerging research has introduced multi-branch convolutional frameworks to Motor Imagery classification, yielding substantial performance gains compared to traditional single-branch baselines. Wu et al. [16] proposed the Multi-Scale Filter Bank Convolutional Neural Network (MSFBCNN) employs temporal convolution layers with multiple scales for extracting diverse features, which are subsequently integrated into a spatial convolution block. Building upon established architectures, Altuwaijri et al. [17] introduced MBEEGNet and MBShallowConvNet. These models extend the standard EEGNet and ShallowConvNet frameworks by integrating a multi-branch design that operates across three distinct kernel sizes. Despite these advancements, the potential of input-dependent dynamic fusion remains largely unexplored in these fixed multi-branch structures in EEG decoding.

Data augmentation strategies for EEG analysis

Data augmentation for multivariate EEG time series is primarily implemented through signal processing techniques that manipulate features within either the time or frequency domains. In the temporal dimension, strategies prioritize geometric transformations and stochastic perturbations to diversify the sample distribution. Conventional methods, such as jittering, scaling, and time warping, modify temporal morphology by injecting Gaussian noise or stretching signal segments, while techniques like channel shuffling recombine data across distinct channels to enhance structural robustness [18]. Regarding the frequency domain, methods typically involve converting signals into time-frequency images and applying specific noise [19]. Recent advancements, such as FreqNoise and FreqMasking, target specific spectral attributes: the former injects complex noise into amplitude and phase information to mimic realistic artifacts, while the latter selectively masks discrete frequency intervals to compel the model to learn more robust representations [20].

Originating from advancements in image synthesis, generative models—specifically Generative Adversarial Networks (GANs) [21] and Diffusion Models [22] have become fundamental frameworks for autonomous distribution modeling. In the context of EEG data augmentation, GANs have seen extensive application, primarily to address class imbalance and enhance signal quality. One study employed GANs to augment a class-imbalanced sleep EEG dataset [23], other researchers utilized GANs to construct a low-sensitivity dataset from a high-sensitivity one, attempting to address the super-resolution reconstruction problem in motor imagery tasks [24]. Despite the effectiveness of GANs in EEG signal processing, recent research has revealed that samples generated by classic GAN models are prone to mode collapse. This issue compromises the similarity between the augmented and original samples [25]. A study employed a diffusion model with a Mixup strategy to generate EEG motor imagery data, by performing data augmentation in both the time and frequency domains, the research achieved favorable results on BCI datasets, offering a potential solution to the scarcity of MI-EEG samples [26]. It was verified that the diffusion model not only boosts the augmented samples’ ability to capture intrinsic features but also prevents GAN-associated mode collapse, effectively mitigating conflicts within ERD/ERS phenomena.

Transfer learning and domain adaptation in EEG decoding task

A major bottleneck in current deep learning methodologies lies in the pronounced heterogeneity of MI-EEG signals across subjects. To address this issue, several Brain-Computer Interface (BCI) studies have adopted transfer learning techniques, which utilize knowledge acquired from a source domain to assist learning in a target domain [27]. Raza et al. [28] introduced a novel framework for detecting and adapting to covariate shifts to alleviate distributional differences in feature spaces. Liu et al. [29] focused on capturing shared features across different subjects to reduce inter-subject variability. Within the broader scope of transfer learning, Domain Adaptation (DA) has emerged as a prominent paradigm specifically designed to align the statistical distributions of source and target domains. Phunruangsakao et al. [30] proposed a few-shot domain adaptation model to reduce inter-subject discrepancies and boost performance on unseen targets. In [31], a dynamic joint domain adaptation framework was put forth, which employed multiple discriminators to effectively reduce distributional shifts between domains. To address cross-subject heterogeneity, [32] proposed a Wasserstein-based adversarial domain adaptation framework that leverages attention mechanisms to extract discriminative features and explicitly minimizes the distributional discrepancy between source and target domains. Huang et al. [33] introduced a shallow Inception domain adaptation framework, which extracts informative deep features from data across multiple subjects to achieve accurate Motor Imagery (MI) recognition. Additionally, some scholars proposed a Deep Domain Adaptation Framework based on Correlation Alignment (DDAF-CORAL) [12] to address the problem of distribution divergence in cross-domain Motor Imagery classification. A study proposed using Multi-Source Self-Selected Joint Domain Adaptation (MSS-JDA) [34] to address the cross-subject problem. The multi-source domain adaptation framework entails building separate branches for each subject’s EEG data. It leverages one-to-one domain alignment to capture domain-specific representations, allowing for inference via a multi-branch architecture [35].

Methods

In this work, we present a unified framework to enhance cross-subject motor imagery classification performance. The framework is composed of two sequential stages:

1. Data Augmentation: We first convert raw 1D EEG signals into 2D time-frequency spectrograms using the Short-Time Fourier Transform (STFT). To address data scarcity, we then use a diffusion probabilistic model built on a U-Net architecture. This network acts as a denoising engine to model the data distribution and generate high-fidelity synthetic samples.

2. AMSC-DANN: Utilizing the augmented dataset, we employ an Adaptive Multi-Scale Convolution with Domain Adversarial Neural Network to extract features and align the source and target domains for effective cross-subject decoding.

Data augmentation

Time-frequency transformation

According to the characteristics of the multivariate time series of motor imagery data, directly using the diffusion model for data augmentation will be affected by certain nonlinear and non-stationary properties [26]. Therefore, converting EEG signals into frequency coefficients for representation has become a processing method in many studies. Meanwhile, diffusion models also have unique advantages in processing image-based data. We first transform the 1D EEG signals into 2D time-frequency representations. Given a raw EEG trial denoted as , where C is the number of channels and T is the time points, we apply the Short-Time Fourier Transform (STFT) to extract spectral-temporal features (Fig. 1).

Fig. 1.

U-Net model for the time–frequency images generation

Given a single-channel EEG signal x [n], the Short-Time Fourier Transform is computed as [36]:

1

where w[n] denotes the window function centered at time index m, and represents the frequency.

Subsequently, we calculate the power spectrogram to capture the energy distribution. To align the data distribution with the input requirements of the diffusion model, the spectrogram is normalized to the range [0, 1] via min-max scaling:

2

where serves as the input image for the subsequent generative module.

In the specific experimental setup, we set the window size to 256 sampling points (approximately 1 s) and the overlap rate to 50% (step size ). This setup achieves a good balance between temporal and frequency resolution. Finally, for each trial, we stack the time-frequency images generated by all channels to form a three-dimensional tensor , where C is the number of channels (22 for Dataset 2a and 3 for Dataset 2b), F is the number of frequency points, and T is the number of time frames. This tensor preserves the spatial structure of the EEG signals and characterizes the energy change patterns of ERD/ERS through time-frequency transformation, which can be directly used as input data for the subsequent Diffusion data augmentation module and AMSC-DANN classification network.

U-net based diffusion model

To address the data scarcity issue, we propose a U-Net [37] based Diffusion Model to capture the distribution of the time-frequency images. The diffusion framework consists of a forward process and a reverse process.Compared with other generative model architectures (CNNs or transformer), the U-Net architecture is particularly advantageous for EEG time-frequency data, as its symmetric encoder-decoder structure with skip connections can effectively preserve fine-grained temporal and spectral details of the EEG signals—details that are critical for distinguishing different motor imagery patterns. Additionally, the U-Net’s downsampling-upsampling mechanism is well-suited to the 2D time-frequency image format derived from EEG signals, enabling efficient extraction of multi-scale features while avoiding information loss during feature transformation. Figure 2. shows a comparison between the EEG time-frequency graph generated by one of the Diffusion models and the original graph. It can be seen that the energy of the key features has been enhanced, and the features of the key parts are clearer.

Fig. 2.

An example of the diffusion data augmentation

Forward Process (Diffusion) The forward process is a Markov chain [38] that gradually adds Gaussian noise [39] to the real data (i.e., ) over T steps. The transition probability at step t is defined as:

3

where is a pre-defined variance schedule. By letting and , we can sample at any arbitrary time step t directly from in closed form:

4

Reverse Process (Denoising): The goal is to learn the reverse distribution to recover the clean image from pure noise . We employ a symmetric U-Net architecture, denoted as , to predict the noise component. The model is trained by minimizing the simplified variational lower bound loss:

5

As shown in Fig. 1, the U-Net architecture comprises an encoder, a decoder, and feature map concatenations bridging the two. The encoder downsamples time-frequency images to generate feature maps, whereas the decoder upsamples these maps to reconstruct the target Gaussian noise. Skip connections are employed to link features from the encoder to the decoder. Both the down-sampling and up-sampling blocks employ a wide residual network for core feature extraction. These noise images are derived as predicted outputs during the backward generation process of the DDPM.

Adaptive multi-scale convolution with domain adversarial neural network

This module consists of an adaptive multi-scale feature extractor and a domain adaptation mechanism, as shown in Fig. 3.

Fig. 3.

Overview of the proposed AMSC-DANN framework. The model integrates multi-scale adaptive feature learning with adversarial domain alignment. Source and target data are projected into a shared bottleneck space to align distributions while preserving discriminative information

Adaptive multi-scale feature extraction

EEG signals often exhibit significant inter-subject variability in spectral-temporal patterns. Standard convolutional networks with fixed kernel sizes may fail to capture features at varying scales. To address this, we propose an AMSC feature extractor, denoted as , the adaptive block of the proposed architecture is shown in Fig. 4. This module dynamically recalibrates the importance of each scale in a data-driven manner, following a three-stage process [40].

Fig. 4.

Schematic of the Adaptive Block. Multi-scale features , , and are summed to form U, from which channel-wise attention weights are generated and used to recalibrate each scale

Given an input tensor , where C is the number of input channels, and F and T represent the frequency and temporal dimensions respectively, we first apply a set of N parallel convolutional transformations , each with a different kernel size k. This yields a set of multi-scale feature maps:

6

where each represents features captured at a specific scale.

To derive guidance for the adaptive selection, the multi-scale features are first aggregated through an element-wise summation, . Subsequently, a global context vector is generated by applying global average pooling across both frequency and temporal dimensions, thereby encoding channel-wise statistics:

7

The global context vector is then fed into a compact attention mechanism to generate scale-specific weights. This is realized by a small multi-layer perceptron (MLP) [41]. First, a bottleneck layer reduces dimensionality to create a compact feature representation :

8

where is a weight matrix with , r is the reduction ratio, denotes Batch Normalization, and is the ReLU activation function. Then, a second linear layer with weights generates attention scores for each scale and channel. These scores are normalized using a softmax function across the scale dimension:

9

where is the attention weight for the k-th scale and c-th channel.

Finally, the output feature map is dynamically synthesized by a weighted sum of the initial multi-scale features, where each feature map is modulated by its corresponding attention vector :

10

where denotes channel-wise multiplication (Hadamard product) [42], with the vector broadcasted across both frequency and temporal dimensions. This adaptive aggregation allows the network to dynamically adjust its effective receptive field based on the input, thus effectively capturing both fine-grained local details and global structural information in a flexible, data-driven manner.

Domain adversarial alignment

To mitigate the distribution shift between the source domain and the target domain , we employ a domain adversarial strategy based on the Domain-Adversarial Neural Network (DANN) [43]. The core idea is to learn a feature representation that is discriminative for the motor imagery classification task but invariant to the subject-specific domain shifts.

The framework consists of a feature extractor , a label classifier , and a domain discriminator D. The domain adaptation is achieved through a two-player minimax game: the domain discriminator D aims to distinguish whether the features come from the source or the target domain, while the feature extractor strives to confuse the discriminator by generating domain-invariant representations. This effectively aligns the global marginal distributions .

The adversarial objective for the domain discriminator is defined as the standard binary cross-entropy loss:

11

During training, we optimize the parameters by seeking a saddle point such that:

12

where is the task classification loss on the source domain, and is a hyperparameter that controls the trade-off between the classification performance and the degree of domain invariance. In our experiments, is gradually increased from 0 to 1 during the training process to suppress early noisy signals. The value of is determined by the following formula:

13

where p is the training progress ranging from 0 to 1, and . This mechanism effectively avoids the interference of domain adversarial loss on the learning of feature representation in the early stage of training. By maximizing , the feature extractor is forced to remove domain-specific information, ensuring that the features extracted from different subjects share a common distribution space.

The classification loss is the standard cross-entropy loss computed on the labeled source domain data:

14

The adversarial loss is formulated as a minimax game, where the feature generator tries to fool the discriminator, while the discriminator tries to identify the domain of the joint representation :

15

Through this optimization, the network effectively learns a domain-invariant feature space that respects class boundaries, ensuring generalization to the target subject.

Experiment and results

Dataset

To evaluate the effectiveness of our proposed framework, we conduct extensive experiments on two public benchmark datasets from BCI Competition IV: Dataset 2a and Dataset 2b. These datasets are widely used for assessing the performance of Motor Imagery (MI) decoding algorithms.Two datasets follow the same timing scheme, as shown in Fig. 5.The dataset descriptions are as follows:

Fig. 5.

Timing scheme of one session under the BCI IV 2a and 2b paradigm [44]

1. BCI Competition IV dataset 2a (BCIIV 2a)

   Dataset 2a consists of EEG data recorded from 9 healthy subjects (A01-A09). The signals were acquired using 22 electrodes positioned according to the international 10-20 system. The sampling rate is 250 Hz. The experimental paradigm involves four classes of motor imagery tasks: left hand, right hand, both feet, and tongue. For each subject, the dataset contains two sessions (training and evaluation) recorded on different days. Each session comprises 288 trials (72 trials per class). The temporal segment of [2, 6] second is considered in our experiments, as shown in Fig. 6.

2. BCI Competition IV dataset 2b (BCIIV 2b)

   Dataset 2b contains EEG data from 9 subjects (B01-B09) performing two classes of MI tasks: left hand and right hand. This dataset focuses on a sparse channel configuration, providing signals from only 3 bipolar channels (C3, Cz, and C4) with a sampling rate of 250 Hz. Each subject participated in 5 sessions, first three sessions were used for training and the rest were used for tests. There are about 400 trials and 320 trials in the training and test sets respectively. And also we use the temporal segment of [3, 7] second in our experiments, as detailed in Fig. 7.

Fig. 6.

Timing scheme of the BCI IV 2a paradigm [44]

Fig. 7.

Timing scheme of the BCI IV 2b paradigm [45]

Data preprocessing

The original data of BCI Competition IV is stored in the Continuous Common Data Format (GDF). When directly used for the training of deep learning models, not only is the dimensionality disaster severe, but also the signal-to-noise ratio is extremely low. There are often multiple artifacts mixed in the data, such as electromyography (EOG), electromyography (EMG), and power frequency interference. Therefore, this study designed a standardized preprocessing method that can extract pure motion intention features and convert them into time-frequency graph representations suitable for processing by convolutional neural networks. First, perform basic processing on the data using Python, and then convert it into time-frequency images through short-time Fast Fourier Transform (STFT) for subsequent data augmentation.

We used the open-source neurophysiological signal analysis library MNE-Python in the Python environment to load and clean the original GDF file. The specific processing procedure is as follows:

1. Data loading and electrode mapping

   Use the mne.io.read.raw.gdf function to read the original file of the BCI Competition IV 2a/2b dataset. According to the international 10-20 system standard, the data channels were mapped to the corresponding scalp spatial positions, the useless EOG channels (the last three channels in Dataset 2a) were eliminated, and the EEG channels were retained for subsequent analysis.

2. Filtering and artifact removal

   The main features of motor imagery are concentrated in rhythm (8–13 Hz) and rhythm (13–30 Hz). To eliminate low-frequency drift and high-frequency electromyographic noise while preserving the complete moving-related frequency band, we apply a fourth-order Butterworth bandpass filter with the frequency range set from 4 to 38 Hz. This frequency band covers theta,alpha,Beta and low-frequency gamma waves, which can fully capture the time-frequency dynamic characteristics of motion intentions.

3. Epoching

   Based on the event markers provided by the dataset, we segmented the consecutive EEG signals into independent Epochs. For Dataset 2a, we extract the data [2,6] seconds after the prompt appears; For Dataset 2b, extract the data of [3,7] seconds. This operation eliminates the early potentials induced by visual cues, ensuring that the model focuses on pure motor imagination mental activities.

4. Normalization

   Due to the significant differences in EEG amplitudes among different subjects or even the same subject at different times, direct input into the network can lead to gradient explosion or convergence difficulties. We adopted the Z-score standardization method to normalize the data of each trial, making its mean 0 and variance 1.

Device and parameters

Our proposed framework was implemented using the PyTorch library in the Python environment. All experiments were conducted on a server equipped with an Intel Xeon(R) Platinum 8352V CPU and an NVIDIA RTX 4090D GPU. Regarding the training strategy, we adopted the Adam optimizer with parameters and . The batch size was set to 64.

For the pre-training of the diffusion model, the learning rate was initialized at and trained for 1,000 epochs to ensure stable generation quality. For the AMSC-DANN classification stage, the model was trained for 500 epochs with an initial learning rate of , which was decayed using a cosine annealing schedule.

Evaluation metrics

We first evaluate the quality and diversity of the generated EEG time-frequency images, we employ the Fréchet Inception Distance (FID) [46]. FID measures the distance between the feature distributions of real data and generated data extracted from a pre-trained Inception-v3 [47] network. A lower FID score indicates that the generated images are more similar to the real images in terms of visual fidelity and diversity.

Mathematically, let and represent the multivariate Gaussian distributions of the feature embeddings for real and generated samples, respectively. The FID is calculated as the Wasserstein-2 distance between these two distributions:

16

where denotes the trace of a matrix, refers to the mean vector, and is the covariance matrix. In our experiments, we compute the FID score using the standard implementation on the entire test set.

To quantitatively evaluate the classification performance, stability, and reliability of our proposed method across different subjects, we employ three standard metrics: Accuracy (Acc), Standard Deviation (Std), and the Kappa coefficient ().

Accuracy measures the proportion of correctly classified samples relative to the total number of test samples. For a specific subject, the accuracy is defined as [48]:

17

where TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives, respectively. In the multi-class scenario of the BCI IV 2a dataset, this is generalized as the ratio of correct predictions to the total sample count. We report both the individual accuracy for each subject and the average accuracy across all subjects to assess overall effectiveness.

Standard Deviation is utilized to evaluate the robustness of the model against inter-subject variability. A lower standard deviation indicates that the model performance is consistent across different individuals, implying better generalization capability. It is calculated as [49]:

18

where N is the total number of subjects (9 for both datasets), denotes the classification accuracy of the i-th subject, and is the average accuracy across all subjects.

Kappa Coefficient () is introduced to evaluate the agreement between the predicted classification and the ground truth, correcting for the possibility of agreement occurring by chance. This metric is particularly robust for handling imbalanced datasets or multi-class problems. The Kappa value is calculated as follows [50]:

19

where represents the observed agreement (identical to accuracy), and represents the hypothetical probability of chance agreement. The value of ranges from − 1 to 1, where indicates perfect agreement, and indicates agreement equivalent to random chance. A higher Kappa value signifies better classifier reliability beyond random guessing.

Result and analysis

We evaluated our method on the BCI IV 2a and 2b datasets and compared it with recent state-of-the-art methods. To comprehensively evaluate the effectiveness of our proposed framework, Table 1 presents a detailed comparison against state-of-the-art methods, including traditional FBCSP, deep learning baselines (ConvNet, EEGNet, Conformer), and domain adaptation approaches (DRDA, DDAN, DAFS), the highest values are highlighted in bold. It is observed that non-adaptive methods struggle with severe domain shifts; for instance, FBCSP relies on handcrafted features that fail to generalize across subjects, while classic deep networks like ConvNet and EEGNet, despite their powerful feature extraction capabilities, exhibit high performance volatility (high standard deviations of 12.7 and 10.5 on dataset 2a) due to the lack of explicit distribution alignment mechanisms. Similarly, the Transformer-based Conformer yields suboptimal results (68.3% on dataset 2a), likely due to overfitting on the limited sample size of BCI datasets. While domain adaptation competitors like DRDA and DAFS mitigate these shifts to some extent, our method achieves superior performance, securing the highest average accuracies of 78.5% and 80.1% on datasets 2a and 2b, respectively. Notably, our model demonstrates exceptional robustness, achieving the lowest standard deviation of 4.5 on dataset 2b compared to 14.8 for Conformer, indicating a stable performance across all individuals. Furthermore, on challenging subjects with low signal-to-noise ratios, such as S02 in dataset 2a, our method significantly outperforms ConvNet (64.1% vs. 55.1%), confirming that our diffusion-augmented adaptation strategy effectively aligns feature distributions and learns invariant representations even under significant inter-subject variability.

Table 1.

Comparisons with state-of-the-art methods on BCI competition datasets IV 2a and 2b (The highest values are highlighted in bold)

Data	Method	S01	S02	S03	S04	S05	S06	S07	S08	S09	Avg.	Std.	Kappa
2a	FBCSP [13]	76.9	57.2	84.9	51.3	52.9	46.5	86.3	84.1	82.2	69.1	14.5	0.59
	ConvNet [14]	76.4	55.1	89.7	74.3	56.6	54.9	92.4	77.9	76.9	72.7	12.7	0.64
	EEGNet [10]	79.9	56.2	89.2	64.9	63.1	58.8	64.3	74.2	77.6	69.8	10.5	0.60
	Conformer [51]	77.2	56.9	80.8	51.5	60.8	62.3	75.1	78.9	70.8	68.3	10.0	0.58
	DRDA [52]	82.5	55.4	87.0	75.7	62.1	57.1	85.6	83.2	82.1	74.5	11.0	0.66
	DDAN [43]	80.8	62.8	77.1	74.1	65.8	64.5	82.7	77.1	77.6	73.2	5.8	0.64
	DAFS [30]	81.9	64.6	88.9	73.6	70.4	56.5	85.3	79.1	81.6	75.7	10.1	0.68
	Ours	86.7	64.1	88.4	75.2	67.3	63.6	88.3	87.3	85.7	78.5	10.3	0.71
2b	FBCSP [13]	68.6	61.7	81.2	90.4	82.3	63.7	84.2	87.2	74.5	76.0	9.2	0.52
	ConvNet [14]	76.7	54.2	53.3	97.1	76.7	81.7	80.5	89.2	80.4	76.7	12.7	0.57
	EEGNet [10]	79.2	58.3	55.8	97.9	65.4	74.2	80.0	91.3	89.2	76.8	13.1	0.61
	Conformer [51]	75.8	58.3	50.4	98.8	76.3	84.2	78.8	92.7	86.3	78.0	14.8	0.59
	DRDA [52]	72.7	63.8	64.6	88.2	80.0	82.4	79.8	81.6	78.9	77.0	7.4	0.62
	DDAN [43]	74.0	71.0	67.8	87.4	82.3	81.0	85.8	80.3	74.1	78.2	5.4	0.67
	DAFS [30]	70.3	71.6	80.9	83.6	80.4	73.5	85.3	87.1	79.6	79.1	6.2	0.66
	Ours	84.5	73.6	75.3	87.3	80.1	73.8	81.2	82.5	82.3	80.1	4.5	0.72

In addition to classification accuracy, we further analyze the reliability of our proposed method using the Kappa coefficient (), as detailed in Table 1. The Kappa metric is crucial as it measures the agreement between predicted and true labels while correcting for the probability of agreement occurring by chance. For the 4-class BCI IV 2a dataset, our method achieves an average value of 0.71. This performance significantly outperforms the baseline Conformer () and the competitive method DAFS (). According to Cohen’s interpretation guidelines, a value in the range of 0.61–0.80 represents substantial agreement. This indicates that the high accuracy of our model is not an artifact of random guessing or class bias, but stems from the model’s genuine discriminative capability across four distinct motor imagery tasks. Similarly, on the BCI IV 2b dataset (2-class), our method maintains superior consistency with a of 0.60, whereas other state-of-the-art methods fluctuate between 0.52 and 0.58. The consistent improvement in Kappa values across both datasets demonstrates that our proposed framework is more robust against inter-subject variability and provides trustworthy predictions, effectively minimizing random classification errors even in challenging EEG environments.

To comprehensively evaluate the fine-grained classification performance of our proposed method, we visualized the confusion matrices for both Dataset 2a (four-class) and Dataset 2b (two-class), as shown in Fig. 8.

Fig. 8.

Confusion matrices of the proposed method on two datasets. The values in the diagonal represent the classification accuracy for each specific class

For Dataset 2a, the model demonstrates robust discriminative ability, with the diagonal accuracy for all four classes ranging from 0.75 to 0.82. Notably, the ‘right hand’ imagery achieves the highest recognition rate (0.82). The most prominent source of error lies in the reciprocal misclassification between contralateral movements, specifically ’left hand’ and ’right hand’ (e.g., 11% of left hand samples are misclassified as right hand). From a neurophysiological perspective, this confusion is a well-known challenge in BCI: motor imagery of either hand primarily triggers Event-Related Desynchronization (ERD) in the bilateral sensorimotor cortices (C3 and C4 channels). The spatial proximity and the occasional bilateral cortical activation during unilateral imagery make their EEG patterns highly overlapping. Furthermore, minor confusions are observed between ’feet’ and ’right hand’ (0.11), which may stem from subject-specific variations in the spatial mapping of cortical source signals to the scalp electrodes.

For Dataset 2b, the binary classification task further isolates this contralateral confusion. The matrix indicates a balanced and strong predictive capability, with the accuracies for ’left hand’ and ’right hand’ reaching 0.78 and 0.82, respectively. The remaining misclassification (around 18%-22%) directly reflects the inherent difficulty in disentangling highly symmetric brain activities. Nevertheless, these stable diagonal values across both datasets confirm that our multi-scale adaptive feature extraction strategy effectively captures the distinct spatial-temporal patterns of motor imagery, maintaining high reliability even under complex cross-subject scenarios.

We also evaluate the optimization efficiency and stability of our proposed method, we visualize the training loss curves on Dataset 2a and Dataset 2b in Fig. 9. As observed, the training process for all comparative methods can be divided into two phases: a rapid descent phase (0–200 epochs) and a fine-tuning phase (200–500 epochs).

Fig. 9.

Training loss convergence analysis on Dataset 2a and Dataset 2b. All models are trained for 500 epochs

In the early stages, the optimization trajectories of different methods are intertwined. However, as training progresses beyond 200 epochs, our method distinguishes itself from competitive baselines (e.g., DAFS and DDAN). Specifically, our approach maintains a continuous downward trend even when other methods begin to plateau, eventually converging to a significantly lower loss value (around 0.05). This demonstrates that our method not only achieves faster convergence but also effectively avoids local minima.

Ablation study

We conducted the following experiments to investigate parameter settings and to demonstrate the practical effectiveness of the diffusion model data augmentation, adaptive multi-convolution, and domain adversarial network in our method.

The diffusion data augmentation

To evaluate the quality and diversity of the generated EEG time-frequency images, we employed the Fréchet Inception Distance (FID) metric. A lower FID score indicates that the distribution of generated images is closer to the real data distribution. We compared our proposed diffusion-based method against classic GAN baselines, specifically DCGAN [53] and WGAN-GP [54]. The results are presented in Table 2. Our method consistently achieves the best performance across both datasets. On Dataset 2a, our model achieves an FID of 22.35, surpassing the strongest baseline (WGAN-GP, 29.64) by a margin of 7.29 points. Similarly, on Dataset 2b, our method records the lowest FID of 18.40, outperforming WGAN-GP (25.12) by approximately 26.7%. The diffusion model significantly outperforms GAN-based methods. While WGAN-GP improves upon DCGAN by stabilizing training, it still struggles to capture the fine-grained high-frequency details typical of EEG spectrograms. Our diffusion approach, through its iterative denoising process, effectively models these complex local textures, resulting in more realistic samples.

Table 2.

Comparison of Fréchet Inception Distance (FID) scores

Method	Model Type	Dataset 2a	Dataset 2b
DCGAN [53]	GAN	40.15	36.80
WGAN-GP [54]	GAN	29.64	25.12
Ours	Diffusion	22.35	18.40

We explored the validity of data augmentation using diffusion models for cross-subject motor imagery, specifically analyzing how augmentation multiples affect accuracy and training time. Table 3 illustrates the effect of varying augmentation ratio. The evidence confirms that the diffusion model significantly enhances classification precision, with accuracy generally rising as the augmentation ratio increases. However, as shown in Table 4, excessively increasing the augmentation ratio provides only marginal gains in accuracy but incurs a substantial cost in training time. As observed, the marginal gain remains high up to the 10 ratio (2.00). However, moving from 10 to 15, the gain suffers a sharp decline to 0.5, indicating that the computational cost outweighs the performance benefits. Therefore, 10 is identified as the optimal saturation point.

Table 3.

Ablation study of diffusion model data augmentation on BCI competition dataset IV 2a (The highest values are highlighted in bold)

Data	Ratio	S01	S02	S03	S04	S05	S06	S07	S08	S09	Avg.
2a	0×	76.2	52.2	78.3	67.9	50.1	51.4	81.5	79.3	77.1	68.3
	1×	81.4	56.2	80.2	66.9	54.6	56.1	86.0	84.2	82.9	72.1
	5×	85.3	65.2	86.5	74.4	66.0	61.3	87.1	85.8	83.3	77.2
	Ours (10×)	86.7	64.1	88.4	75.2	67.3	63.6	88.3	87.3	85.7	78.5
	15×	87.2	65.0	86.9	75.8	66.8	64.4	88.6	87.0	87.4	78.7
	30×	88.1	63.7	88.2	76.4	68.5	62.3	89.2	87.2	86.8	78.9

Table 4.

Quantitative trade-off analysis between accuracy and training cost

Ratio	Accuracy (%)	Time (h)	Marginal gain
0	68.5	0.1	–
1	72.2	0.2	37.00
5	77.3	0.5	17.00
10	78.5	1.1	2.00
15	78.7	1.5	0.50
30	78.9	2.8	0.15

The marginal gain represents the accuracy improvement per unit of training time ().

The adaptive multi-scale convolution

To validate the effectiveness of our adaptive multi-scale convolution module, we compared its performance against baseline models with fixed receptive fields. Table 5 presents the accuracy of our full model versus variants that utilize only a single kernel size (3 × 3) and a static fusion of multiple kernels (3 × 3, 5 × 5, and 7 × 7). The “single 3 × 3 kernel” baseline is limited by a restricted receptive field, resulting in the lowest average accuracies. In contrast, the “Static multi-kernel” architecture structurally expands the feature view through parallel convolutions, yielding a substantial performance leap of over 6% on both datasets. This confirms the structural necessity of multi-granular representation for handling non-stationary EEG signals. Building on this, our “Adaptive” approach introduces a dynamic attention mechanism to overcome the limitations of static weighting. Consequently, the performance is further refined to 78.5 and 80.1%, with superior results across almost all individual subjects. This indicates that the structural flexibility provided by the adaptive mechanism is crucial for maximizing model potential and robustness against individual variability.

Table 5.

Ablation study of the adaptive multi-scale convolution module on BCI competition dataset IV 2a and 2b (The highest values are highlighted in bold)

Data	Conv. Module	S01	S02	S03	S04	S05	S06	S07	S08	S09	Avg.
2a	Single 3x3 kernel	80.2	56.1	74.5	70.3	58.0	58.5	80.1	82.0	76.5	70.6
	Static multi-kernel	86.5	62.8	85.1	73.6	65.2	61.3	86.4	85.5	83.9	76.7
	Ours	86.7	64.1	88.4	75.2	67.3	63.6	88.3	87.3	85.7	78.5
2b	Single 3x3 kernel	73.5	62.1	67.7	78.3	72.2	64.2	75.4	73.9	74.6	71.3
	Static multi-kernel	81.2	73.9	72.3	84.2	79.4	70.5	78.7	80.4	79.9	77.8
	Ours	84.5	73.6	75.3	87.3	80.1	73.8	81.2	82.5	82.3	80.1

We also visualize the learned feature distributions using t-SNE. As shown in Fig. 10, the features extracted by the AMSC module exhibit distinct cluster structures with high intra-class compactness and clear inter-class separability. In contrast, the baseline model using a fixed single-scale convolution, results in blurred decision boundaries and overlapping between categories. This comparison demonstrates that the design of AMSC effectively captures multi-granular temporal-spectral patterns.

Fig. 10.

t-SNE feature distribution visualization of subject A05 and B04. (A) and (B) show the 4-class feature embeddings with and without AMSC, respectively. (C) and (D) show the 2-class embeddings on BCI-IV 2b with and without AMSC

The domain adversarial network

To evaluate the effectiveness of the proposed domain adaptation (DA) strategy, we conducted an ablation study on the BCI Competition IV 2a and 2b datasets (Table 6). Without DA (“No DA”), the model suffers from severe performance degradation, averaging only 69.9% and 71.0% accuracy. This confirms that raw EEG signals contain highly subject-specific variations, causing the classifier to overfit the source domain if not explicitly aligned.

Table 6.

Ablation study of domain adaptation strategies on BCI Competition IV 2a and 2b datasets (The highest values are highlighted in bold)

Data	DA Strategy	S01	S02	S03	S04	S05	S06	S07	S08	S09	Avg.	Std.
2a	No DA	74.5	57.0	78.0	69.5	59.3	56.7	80.2	78.5	75.2	69.9	9.1
	MMD	77.2	60.2	82.5	69.4	64.7	58.1	80.9	82.4	79.1	72.7	9.2
	CORAL	79.1	62.5	83.1	72.3	64.5	62.0	82.8	85.8	80.6	74.7	9.1
	Ours	86.7	64.1	88.4	75.2	67.3	63.6	88.3	87.3	85.7	78.5	10.3
2b	No DA	68.0	64.8	74.1	73.2	68.2	64.3	77.1	76.5	73.0	71.0	4.4
	MMD	75.3	66.2	73.4	81.8	74.1	68.6	77.6	79.1	74.2	74.5	4.6
	CORAL	75.0	68.9	77.1	81.5	75.8	72.3	76.3	78.8	76.0	75.7	3.6
	Ours	84.5	73.6	75.3	87.3	80.1	73.8	81.2	82.5	82.3	80.1	4.5

Traditional alignment methods like MMD and CORAL alleviate this by minimizing empirical statistical distances (e.g., mean embeddings or covariance matrices) between domains, bringing the average accuracy up to approximately 74.8% (2a) and 75.7% (2b). However, these methods rely on shallow statistical moments and often struggle to capture the complex, non-linear distribution shifts inherent in heterogeneous EEG data.

In contrast, our proposed method consistently outperforms these baselines, achieving the highest average accuracies of 78.5% (2a) and 80.1% (2b). This superior performance stems from our specific architectural design. As illustrated in our framework, we employ an adversarial adaptation strategy based on a Domain-Adversarial Neural Network (DANN). By introducing a Gradient Reversal Layer (GRL) and a Domain Discriminator, the feature extractor actively learns to deceive the discriminator, dynamically mapping both source and target data into a shared, domain-invariant latent space at the bottleneck layer. Furthermore, the integration of multi-scale convolutions ( 33,55,77) and the “Adaptive block” ensures that fine-grained, task-relevant motor imagery patterns are preserved during the adversarial alignment process.

The advantage of this structural design is particularly evident in subjects with severe distribution shifts, where conventional statistical matching often fails. For instance, on subject S01 (+12.2% on 2a; +16.5% on 2b) and S04 (+14.1% on 2b), our method yields dramatic improvements over the “No DA” baseline. By effectively filtering out subject-specific noise while preserving discriminative semantic features, our adversarial framework successfully bridges the inter-subject gap, enabling robust cross-subject decoding.

Discussion

Advantages of our proposed work

Impact of Unet-based diffusion data augmentation

The scarcity of labeled EEG data and the high inter-subject variability are the main obstacles in BCI calibration. Our results, particularly the significantly reduced standard deviation, demonstrate the effectiveness of our diffusion-based augmentation strategy. Unlike traditional geometric augmentations (cropping, noise injection) that may disrupt the temporal continuity of EEG signals, or GAN-based methods that often suffer from mode collapse, the diffusion model learns the intrinsic data manifold to generate diverse and high-fidelity samples.

By expanding the training set with these generated samples, we effectively bridge the distribution gap between the source and target domains. This “dense” feature space allows the classifier to learn smoother decision boundaries, thereby improving robustness against noise and outliers. This is evident in the substantial performance boost on difficult subjects, where the generated data supplemented the sparse features of the original low-quality signals.

Efficacy of the AMSC-DANN architecture

While data augmentation provides high-quality inputs, the specific design of our AMSC-DANN architecture ensures that the most discriminative features are extracted. A critical innovation here is the Adaptive Multi-Scale Convolution (AMSC) based on the Selective Kernel (SK) network mechanism [40]. Traditional multi-scale methods fuse features from fixed kernel sizes statically. In contrast, our AMSC module dynamically adjusts the receptive field size based on the input signal characteristics. EEG signals exhibit significant non-stationarity; for instance, the duration and frequency of motor imagery patterns (ERD/ERS) can vary dramatically across subjects and trials.

Through the attention-guided selection mechanism, the AMSC module can automatically assign higher weights to the most relevant kernel branch. This means the network adaptively decides whether to focus on short-term transient details or long-term rhythmic dependencies for each specific input sample. Combined with the domain adversarial alignment, this adaptive capability allows our model to capture robust, subject-invariant representations that are often missed by fixed-scale architectures, directly contributing to the state-of-the-art accuracy achieved in our experiments.

Limitations and future work

While our proposed method achieves competitive overall performance, the confusion matrices (Fig. 8) reveal a persistent bottleneck: the mutual misclassification between highly symmetric tasks, particularly ’left hand’ and ’right hand’. To further alleviate this confusion in future work, several advanced strategies can be explored:

1. Spatial attention mechanisms: Although our multi-scale convolutions capture broad spatial features, integrating a dynamic spatial attention module could help the network strictly focus on the discriminative activation differences between the C3 (left hemisphere) and C4 (right hemisphere) regions, thereby suppressing ipsilateral noise.

2. Supervised contrastive learning: The current cross-entropy loss focuses on absolute class probabilities. By introducing a supervised contrastive loss, the network can be explicitly forced to push the latent representations of ’left hand’ and ’right hand’ further apart in the feature space, maximizing inter-class margins for these hard-negative pairs.

3. Brain functional connectivity: Traditional CNNs extract localized grid-based features. Incorporating Graph Neural Networks (GNNs) to model the phase-locking value (PLV) or coherence between different brain regions could capture the asymmetrical functional connectivity routing during unilateral motor imagery, providing a critical complementary feature to distinguish contralateral movements.

Despite the performance, our framework entails higher computational costs compared to lightweight models like EEGNet, primarily due to the iterative denoising process of the diffusion model. Training time is significantly increased as the model learns to reconstruct the complex EEG manifold. Future work will focus on exploring accelerated sampling methods, such as DDIM [55] or consistency models, to further improve the training efficiency and facilitate rapid calibration for new users.

Conclusion

In this paper, we addressed the critical challenges of data scarcity and inter-subject variability in EEG-based motor imagery classification. We proposed a comprehensive framework that synergizes generative data augmentation with a domain-adversarial deep learning architecture. Specifically, we leveraged a diffusion probabilistic model to synthesize high-fidelity EEG samples, effectively expanding the training distribution and mitigating the risk of overfitting. Furthermore, the proposed AMSC-DANN architecture integrates Adaptive Multi-Scale Convolutions to capture dynamic temporal-spectral features at varying granularities, coupled with a DANN to enforce feature alignment across subjects. Extensive empirical evaluations on the BCI Competition IV datasets demonstrate that our method outperforms state-of-the-art baselines, achieving classification accuracies of 78.5% on dataset 2a and 80.1% on dataset 2b, while maintaining relatively low standard deviations. Ablation studies further verified the distinct contribution of each module: the adaptive mechanism is crucial for extracting discriminative patterns from non-stationary signals, while the adversarial strategy effectively reduces the distribution divergence between source and target domains. Future work will focus on refining the core components. We aim to enhance the diffusion model with efficient sampling strategies and advanced conditioning to further improve signal fidelity. Additionally, we plan to upgrade the domain adaptation module by incorporating class-conditional or metric-based strategies, enabling fine-grained feature alignment to minimize negative transfer and boost generalization.

Author contributions

J. Zhang proposed the overall framework, implemented the neural network models, and drafted the original manuscript. H. Zhang and Y. Yang provided theoretical guidance and constructive suggestions, and guided the revision of the manuscript. H. Zhang supervised the study and secured the funding. All authors reviewed and approved the final manuscript.

Funding

This work was supported in part by the Humanity and Social Science Foundation of the Ministry of Education of China under Grant 21A13022003 and in part by from the Natural Science Foundation of Zhejiang Province (Grant No. LMS26F020033).

Data availability

Publicly available datasets were analyzed in this study. This data can be found here: http://www.bbci.de/competition/iv/. Specifically, we used Dataset 2a and Dataset 2b from the BCI Competition IV. Further details regarding the data acquisition and experimental paradigm can be found in the original competition description and related publications.

Declarations

Competing interests

The authors declare no conflict of interest.