NeuroLightNet: A Lightweight Attention-Driven Deep Learning Framework for Adaptive Electroencephalogram Signal Interpretation in Brain–Computer Interfaces

The remainder of the paper is organized as follows: Section 2 presents a comprehensive review of related works, highlighting recent developments in deep learning-based EEG classification, attention mechanisms, continual learning, and explainability in BCIs. Section 3 details the proposed methodology, elaborating on the design and components of the NeuroLightNet framework, including the NeuroCRAMNet architecture and the integration of attention and continual learning modules. Section 4 describes the experimental setup, including dataset specifications, preprocessing techniques, evaluation metrics, and deployment considerations. Section 5 discusses the results and provides a comparative analysis of the proposed method against baseline models using multiple evaluation criteria and visualizations. Section 6 concludes the paper with a summary of findings and outlines directions for future research, including multi-modal extensions and adaptive personalization strategies. In contrast to traditional CNN-GRU or attention-based EEG classification systems, the NeuroLightNet framework suggests a single architecture that can be used to tackle the idea of accuracy, flexibility, interpretability, and the efficiency of its deployment. What is new about this work is not in the separate components, but in the synergistic combination and task-specific optimization of these components to EEG-based BCI applications. Particularly, NeuroCrackMe includes a two-stage spatial-temporal attention system to improve the discriminative features learning, implements the EWC method to continually learn across sessions and subjects, and introduces explainable AI methods, such as SHAP and Grad-Cam, directly into the model. Moreover, the framework is specifically designed to deploy depthwise separable convolutions, quantization, and Open Neural Network Exchange (ONNX)-based implementation. This is a multi-objective architecture that makes NeuroLightNet different than current CNN-GRU and attention-based EEG models, a full-fledged and scalable solution to adaptive BCI systems. To overcome the constraints of the currently available EEG-based BCI models, this paper will present NeuroLightNet, a lightweight and versatile framework of the deep learning network that is specifically built to interpret EEG signals effectively. The suggested NeuroCRAMNet model combines depthwise-separable convolution (that enables the efficient extraction of spatial features), GRU-based temporal modeling, and dual-stage spatial-temporal attention to improve discriminative learning. The framework, further, has EWC of session- and subject-wise continuous learning and embedded explainable AI methods (SHAP and Grad-CAM) to enhance model interpretability. This single design allows high classification accuracy, computational efficiency, and possible deployment in real-time applications, unlike the current EEG classification frameworks.

The proposed work introduces NeuroLightNet, a lightweight deep learning framework for accurate and adaptive EEG signal interpretation in BCIs. It leverages NeuroCRAMNet, which integrates depthwise-separable convolutions, GRU-based temporal modeling, and dual-stage attention mechanisms for efficient spatiotemporal feature extraction. A continual learning module based on EWC ensures adaptability across sessions and subjects. Explainable AI techniques (SHAP, Grad-CAM) are embedded for model transparency and trust. The framework is optimized for real-time edge deployment, achieving high accuracy with minimal latency and memory footprint.

3.1 Dataset description

Two benchmark datasets were employed to validate the proposed NeuroLightNet framework. The BCI Competition IV-2a dataset comprises EEG recordings from 9 subjects performing four-class motor imagery tasks (left hand, right hand, foot, tongue) using 22 channels at 250 Hz. Each subject participated in two sessions recorded on different days. The PhysioNet Motor Imagery EEG dataset includes recordings from 109 subjects executing or imagining hand and foot movements across 64 EEG channels at 160 Hz. This dataset provides a broader demographic with varying session protocols. Together, these datasets enable rigorous evaluation across spatial, temporal, and inter-subject variations in EEG signals, as shown in Table 2.

Table 2. Dataset parameters of the Electroencephalogram (EEG) datasets used

3.2 Preprocessing

Preprocessing is a crucial step in EEG-based BCI systems to enhance signal quality and ensure meaningful feature extraction. Raw EEG signals are often contaminated with noise, artifacts, and irrelevant frequency components, which can impair model performance. The goal of preprocessing is to clean, standardize, and structure the data for optimal input to the deep learning framework. The following steps outline a rigorous pipeline tailored to the BCI Competition IV-2a and PhysioNet Motor Imagery, and it is shown in Figure 2.

Figure 2. Preprocessing techniques

3.2.1. Bandpass filtering

To isolate relevant EEG frequency bands (e.g., μ: 8–12 Hz and β: 13–30 Hz), a Butterworth bandpass filter is applied. Let x(t) be the raw EEG signal. The bandpass filtered signal x_f(t) is given by:

${{x}_{f}}\left( t \right)={{F}_{BP}}\left\{ x\left( t \right) \right\}$                   (1)

where, ${{F}_{BP}}$ represents the bandpass filter operator. For an n-order Butterworth filter:

$H\left( \omega  \right)=\frac{1}{\sqrt{1+{{\left( \frac{\omega }{{{\omega }_{c}}} \right)}^{2n}}}}$                  (2)

where, ${{\omega }_{c}}$ is the cutoff frequency, and n is the filter order (typically 4 or 5).

3.2.2 Artifact removal using independent component analysis

To eliminate ocular and muscular artifacts, independent component analysis (ICA) decomposes the EEG into statistically independent components.

Let X ∈ R^C×T be the recorded EEG matrix with C channels and T time samples. ICA attempts to find an unmixing matrix W such that:

$S=WX$                      (3)

where, S contains statistically independent sources. Artifact components are identified and zeroed out, then the cleaned signal is reconstructed:

${{X}_{clean}}={{W}^{-1}}{{S}_{clean}}$                           (4)

3.2.3 Epoch extraction

EEG data are segmented into fixed-length epochs around stimulus onset events. Let the onset time be t_s, the pre-stimulus window T_p, and the post-stimulus window T_q. The epoch E_i(t) for the i_th trial is:

${{E}_{i}}\left( t \right)=x\left( t \right),t\in \left[ {{t}_{s}}-{{T}_{p}},{{t}_{s}}+{{T}_{q}} \right]$                (5)

Typically, ${{T}_{p}}$ = 0.5 sec and ${{T}_{q}}$ = 2 sec for motor imagery.

3.2.4 Normalization (Z-score scaling)

To ensure uniform channel contribution and stabilize training, z-score normalization is applied across each channel. For a channel signal x_c:

$x_{c}^{norm}=\frac{{{x}_{c}}-{{\mu }_{c}}}{{{\sigma }_{c}}}$                      (6)

where, ${{\mu }_{c}}$ and ${{\sigma }_{c}}$ are the mean and standard deviation of channel c.

3.2.5 Spatial filtering using Common Spatial Pattern

CSP enhances class-discriminative spatial patterns. Let Σ₁ and Σ₂ be the covariance matrices for two classes. The CSP projection matrix W_CSP solves:

${{W}_{CSP}}=arg\underset{w}{\mathop{\max }}\,\frac{{{W}^{T}}\mathop{\sum }_{1}W}{{{W}^{T}}(\mathop{\sum }_{1}+\mathop{\sum }_{1})~W~}$               (7)

The transformed signal:

$Z=W_{CSP}^{T}X$ selects components with maximal variance differences between classes.

3.2.6 Temporal Smoothing (Moving average or Savitzky-Golay filter)

To reduce temporal noise and highlight meaningful trends in EEG sequences, a smoothing operator S is applied.

$\tilde{x}\left( t \right)=\frac{1}{2k+1}\underset{i=-k}{\overset{k}{\mathop \sum }}\,x\left( t+i \right)$             (8)

where, k is the window size (e.g., k = 5).

$\tilde{x}\left( t \right)=\underset{i=-k}{\overset{k}{\mathop \sum }}\,{{c}_{i}}x\left( t+i \right)$ where ${{c}_{i}}$ is the polynomial fit coefficients for local smoothing.

3.2.7 Downsampling

To reduce computational load, the signal is downsampled from the original sampling rate f_s to a lower rate f_d.

${{x}_{d}}\left[ n \right]=x\left[ n\cdot D \right],D=\frac{{{f}_{s}}}{{{f}_{d}}}$               (9)

This preserves essential dynamics while reducing sequence length.

3.3 Proposed methodology

The core of the NeuroLightNet framework is a novel architecture called NeuroCRAMNet (Compressed Recurrent Attention Model), designed to efficiently extract spatial-temporal features from EEG signals while enabling attention-driven focus, continual adaptability, and edge-ready deployment. The architecture integrates four primary functional modules: compressed spatial encoders, temporal attention blocks, continual learning units, and explainable AI heads. Each of these components is developed with considerations of low-latency inference, robustness to inter-session variability, and interpretability—features critical to the deployment of EEG-based BCIs in real-world applications. In order to make it reproducible and allow future research, the detailed setting of the suggested NeuroCRAMNet architecture is displayed in this subsection.

The spatial feature extraction module consists of two depthwise-separable convolutional layers. The initial layer consists of 32 filters with the usage of (1 × 3) kernel size, batch normalization, and ReLU activation. The second layer involves 64 filters with a kernel (1 × 5) so as to obtain wider spatial dependencies between EEG channels. The layers are both followed by a dropout rate of 0.3 in order to avoid overfitting. The temporal modelling part makes use of a GRU layer with 128 hidden units, which allows learning sequential dependencies in EEGs with computational efficiency. GRU output is then subjected to a temporal self-attention (TSA) mechanism comprising 4 attention heads, whose key/query dimension is 32. In the case of spatial attention, a multi-head attention module consisting of 4 attention heads is utilized following the convolutional encoder to increase feature representations on a channel basis. There are residual connections and layer normalization, which are added to stabilize training and enhance convergence. Elements of the classification phase are a fully connected, dense layer of 256 neurons and a multi-class EEG signal classification using a softmax activation function. In order to increase generalization performance, the L2 regularization of a 1e-4 coefficient is added to all trainable parameters. The amount of all trainable parameters of the proposed NeuroCRAMNet model is about 1.18 million, which is much less than the traditional CNN-GRU architectures and yet generates better classification results. The most important architectural parameters of the proposed model are summarized in Table 3.

Table 3. Dataset summary

3.3.1 Spatial feature encoding via depthwise-separable convolution and channel attention

To capture spatial patterns across EEG channels, the input EEG matrix X ∈ RC × T, where C is the number of channels and T is the number of temporal samples, is first reshaped to fit a convolutional input format. Depthwise-separable convolution is employed as an efficient alternative to traditional convolution. This operation decomposes standard convolution into two steps: depthwise and pointwise convolution, and it is shown in Figure 3. In the depthwise step, a single convolutional filter is applied per input channel, defined as:

${{X}_{d}}\left( c,t \right)=\underset{k=1}{\overset{K}{\mathop \sum }}\,{{w}_{c}}\left( k \right)\cdot X\left( c,t+k \right)$                (10)

where, ${{w}_{c}}$ represents the depthwise kernel for the cth channel, and K is the kernel size. The resulting output is then passed through a pointwise convolution, which performs a 1 × 11 convolution across all channels:

${{X}_{p}}\left( t \right)=\underset{c=1}{\overset{C}{\mathop \sum }}\,{{V}_{c}}\cdot {{X}_{d}}\left( c,t \right),$                  (11)

where, ${{V}_{c}}$ are the learned weights for inter-channel integration. This design drastically reduces computational complexity from O(C + K), making it suitable for edge deployment.

To enhance the spatial discriminative capability, a lightweight multi-head attention mechanism is introduced after the convolution. The attention module learns to focus on informative channels by computing attention scores using the scaled dot-product:

$Attention\left( Q,K,V \right)=softmax\left( \frac{Q{{K}^{T}}}{\sqrt{{{d}_{k}}}} \right)V$                   (12)

where, the query Q, key K, and value V matrices are derived from linear projections of the input: Q = XW^Q, K = XW^K, and V = XW^V. The attention heads are concatenated and added to the original input via a residual connection, ensuring gradient flow and learning stability. Spatial Feature Encoding via Depthwise-Separable Convolution and Channel Attention is shown in Figure 3.

Figure 3. Spatial feature encoding via depthwise-separable convolution and channel attention

3.3.2 Temporal modeling using gated recurrent units and self-attention

Following spatial encoding, the signal is processed through a temporal learning unit to extract meaningful sequential dependencies. A GRU is used for this purpose due to its lower parameter count and competitive performance in time-series modelling, and it is shown in Figure 4.

Figure 4. Temporal modeling using gated recurrent units (GRUs) and self-attention

The GRU computes its hidden state at time t using the following update equations:

${{z}_{t}}=\sigma \left( {{W}_{z}}{{x}_{t}}+{{U}_{z}}{{h}_{t-1}} \right),{{r}_{t}}=\sigma \left( {{W}_{r}}{{x}_{t}}+{{U}_{r}}{{h}_{t-1}} \right)$                  (13)

${{\tilde{h}}_{t}}=\tanh \left( {{W}_{h}}{{x}_{t}}+{{U}_{h}}\left( {{r}_{t}}\odot {{h}_{t-1}} \right) \right)$                         (14)

${{h}_{t}}=\left( 1-{{z}_{t}} \right)\odot {{h}_{t-1}}+{{z}_{t}}\odot {{\tilde{h}}_{t}}$                         (15)

where, z_t and r_t represent the update and reset gates, respectively, and ⊙ denotes the Hadamard product.

To further refine the temporal features, a self-attention mechanism is applied across the hidden states generated by the GRU. Given a sequence of hidden states H = [h₁,h₂,...,h_T] attention weights α_t are computed as:

${{\alpha }_{t}}=\frac{exp\left( {{u}^{T}}tanh\left( {{W}_{a}}{{h}_{t}} \right) \right)}{\mathop{\sum }_{i=1}^{T}exp\left( {{u}^{T}}tanh\left( {{W}_{a}}{{h}_{i}} \right) \right)}$                 (16)

and the context vector is formed as a weighted sum of hidden states:

${{h}_{context}}=\underset{t=1}{\overset{T}{\mathop \sum }}\,{{\alpha }_{t}}{{h}_{t}}$                (17)

This process allows the network to dynamically focus on salient temporal intervals, such as P300 spikes or motor imagery activation periods.

3.3.3 Continual learning through elastic weight consolidation

To address the problem of inter-session and inter-subject variability, continual learning is embedded within the training process using EWC. This regularization-based approach prevents catastrophic forgetting when adapting the model to new data distributions, as shown in Figure 5.

Figure 5. Continual learning through elastic weight consolidation (EWC)

After training on a previous task A, a diagonal approximation of the Fisher Information Matrix FFF is computed to estimate the importance of each parameter θ_i. The model is then fine-tuned on the new task BBB with the loss function augmented as:

${{L}_{total}}={{L}_{B}}\left( \theta  \right)+\frac{\lambda }{2}\underset{i}{\mathop \sum }\,{{F}_{i}}{{\left( {{\theta }_{i}}-\theta _{i}^{*} \right)}^{2}}$              (18)

where, $\theta _{i}^{*}~$denotes the learned parameters from task A, and λ controls the regularization strength. The Fisher Information is estimated using:

$F_i=E\left[\left(\frac{\partial \log _{\mathrm{p}}(\mathrm{y}\mid\mathrm{x}, \theta)}{\partial \theta_i}\right)\right]$              (19)

Approximated over the training data using the model’s output probabilities. This mechanism enables the network to learn from new subjects or sessions while retaining prior knowledge.

3.3.4 Explainability integration using SHapley Additive exPlanations and Gradient-weighted Class Activation Mapping

To enhance model interpretability and user trust, explainable AI components are integrated into the output layer. SHAP (Shapley Additive Explanations) is used to compute the contribution of each input feature to the final decision, and it is shown in Figure 6. The SHAP value ϕ_i for the i_th feature is defined as:

${{\phi }_{i}}=\underset{S\subseteq F\setminus \left\{ i \right\}}{\mathop \sum }\,\frac{\mid S\mid !\left( \mid F\mid -\mid S\mid -1 \right)!}{\mid F\mid !}\left[ f\left( S\cup \left\{ i \right\} \right)-f\left( S \right) \right]$                (20)

where, F is the feature set, and f(S) is the model’s output when only features in subset S are present. Additionally, Grad-CAM is applied to identify which spatial-temporal regions of the input most strongly influenced the model’s prediction. Grad-CAM computes the class-discriminative localization map as:

$\alpha _{k}^{c}=\frac{1}{z}\underset{i}{\mathop \sum }\,\underset{j}{\mathop \sum }\,\frac{\partial {{y}^{c}}}{\partial A{{_{ij}^{k}}^{{}}}}$                 (21)

$GradCA{{M}^{c}}=ReLU\left( \underset{k}{\mathop \sum }\,\alpha _{k}^{c}{{A}^{k}} \right)$                      (22)

where, A^k is the kth activation map, y^c is the class score, and Z is the normalization factor. These tools provide insight into which regions of the EEG signal (in space and time) were critical for each classification decision.

Figure 6. Explainability integration using SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM)

3.3.5 Edge deployment via quantization and Open Neural Network Exchange optimization

One of the distinguishing features of NeuroCRAMNet is its readiness for deployment in real-time, low-power environments such as embedded systems, wearable devices, or neuroprosthetic controllers. To meet the constraints of latency, memory, and energy efficiency, we implement post-training model compression techniques, including parameter pruning, weight quantization, and ONNX conversion for hardware compatibility.

The primary technique employed is 8-bit integer quantization, which reduces the model's floating-point precision to a lower bit-width representation without significantly affecting accuracy. Consider a 32-bit floating-point weight θ_fp32; quantization maps this to an 8-bit integer θ_int8 using a uniform scaling factor s, defined as:

${{\theta }_{int8}}=round\left( \frac{{{\theta }_{fp32}}}{s} \right)),$with $s=\frac{max\left( {{\theta }_{fp32}} \right)-min\left( {{\theta }_{fp32}} \right)}{{{2}^{n}}-1}$                 (23)

where, n = 8 and s serves to preserve the dynamic range of the original weights. During inference, dequantization is performed to approximate the original weights:

${{\theta }_{fp32}}\approx s\cdot {{\theta }_{int8}}.~$                                  (24)

This operation significantly reduces the memory footprint and accelerates matrix multiplications on integer-only hardware accelerators such as ARM Cortex-M microcontrollers or edge AI chips like NVIDIA Jetson Nano and Intel Movidius.

In addition to quantization, model pruning is employed to remove redundant or low-importance weights. This is guided by magnitude-based criteria where weights θ_i with absolute values below a threshold δ are zeroed out:

$\theta _{i}^{'}=\left\{ \begin{matrix}   0,~if~\left| {{\theta }_{i}} \right|<~\delta   \\   {{\theta }_{i}},~Otherwise  \\ \end{matrix} \right.$                     (25)

Pruning not only reduces parameter count but also sparsifies the model, enabling further computational speedup through sparse matrix optimizations. Finally, the trained model is exported in ONNX format, which facilitates cross-platform compatibility and seamless deployment on a variety of edge AI frameworks such as TensorRT, TFLite, and EdgeTPU. Empirical benchmarks show that the fully optimized model achieves inference latencies under 20 milliseconds per trial with a total memory usage below 8 MB, satisfying the stringent real-time constraints of practical BCI applications. By integrating compression and deployment capabilities directly into the network design, NeuroCRAMNet bridges the gap between deep learning research and practical neurotechnology systems, paving the way for efficient, interpretable, and mobile brain–computer interaction.