EEG Decoding Using CNN and LSTM Network

This study introduces a hybrid deep-learning architecture that integrates a convolutional neural network (CNN) with a bidirectional long short-term memory (bi-LSTM) network for motor imagery (MI) brain–computer interface (BCI) EEG decoding. The CNN is used to learn high-level spatial and temporal representations directly from raw MI-EEG recordings, whereas the bi-LSTM models temporal dependencies and relationships among the extracted features. The proposed approach is evaluated using both a publicly available dataset and a privately acquired dataset obtained with an EEG acquisition system. The experimental results indicate that the CNN&bi-LSTM architecture provides robust performance for both two- and three-class motor-imagery classification and demonstrates promising subject-independent decoding capability across the evaluated methods.

The paper addresses the challenge that MI-EEG signals are affected by multiple sources of noise and generally exhibit a low signal-to-noise ratio. Conventional approaches such as Linear Discriminant Analysis (LDA), Support Vector Machines (SVMs), and Naive Bayes (NB) have limitations for nonlinear, noisy, and correlated EEG data. Deep neural networks can learn latent structures directly from raw measurements, reducing reliance on manually designed feature-selection and feature-extraction stages. However, conventional deep-learning pipelines based on individual CNNs or RNNs still require substantial effort for model training, parameter optimization, and EEG feature engineering.

The proposed framework consists of a preprocessing stage followed by a convolutional neural network and a bidirectional LSTM module. The preprocessing includes referencing using the vertex electrode (Cz), electrode selection (P3, P4, C3, C4, O1, O2, Pz, Fz, and Cz), wavelet-based denoising, and signal filtering within the 8–23 Hz frequency range covering µ and beta rhythms. The CNN employs eight spatial filters with a convolution kernel of size [36 × 1], producing eight feature representations after convolution. Batch normalization is incorporated to improve generalization and accelerate optimization. The bidirectional LSTM processes the sequence in both directions, incorporating information from preceding as well as subsequent observations, which improves the representation of temporal context and provides greater robustness during training.

The experimental results show that the proposed CNN&bi-LSTM model achieves a mean classification accuracy of 85.4% on the private dataset D1, compared with 69.5% for LDA, 73.0% for SVM, 75.9% for CNN, and 79.5% for RNN. The proposed model achieved its highest subject-specific accuracy of 92.5% for subject C, followed by 91.4% for subject E and 90.3% for subject F. On the public datasets, the CNN&bi-LSTM configuration achieves 81.80%, 87.97%, and 85.32% on D2, D3, and D4, respectively, outperforming ST-CNN, RNN, and LSTM models. The results indicate that combining convolutional feature extraction with bidirectional temporal modeling is more effective than simply increasing the number of recurrent parameters or layers.

The paper concludes that the CNN&bi-LSTM architecture provides a unified mechanism for transforming lower-level EEG measurements into higher-level representations that incorporate both temporal and frequency-related information. The experimental results demonstrate that the method can effectively emphasize informative components of MI-EEG signals and achieve competitive classification performance across the evaluated datasets and subjects. The use of bidirectional recurrent modeling represents a potential alternative to conventional spectral-feature-based EEG analysis and may support future research and practical applications of MI-EEG-based rehabilitation systems.

Improvements to AI Systems:

1. Hybrid Spatial-Temporal Feature Extraction: Integrate a CNN front-end (with spatial filters like the [36×1] kernels) to automatically learn high-level spatial patterns from raw EEG, followed by a bidirectional LSTM to model temporal dependencies in both forward and backward directions. This replaces manual feature engineering and improves robustness to low signal-to-noise ratios.

2. Subject-Independent Decoding via Bidirectional Context: Use the bi-LSTM’s ability to incorporate future and past context to generalize across different subjects without retraining, enabling plug-and-play BCI systems for new users.

3. Noise-Resilient Preprocessing Pipeline: Adopt the specific preprocessing chain—Cz referencing, electrode selection (P3, P4, C3, C4, O1, O2, Pz, Fz, Cz), wavelet denoising, and 8–23 Hz bandpass filtering—to suppress artifacts and enhance µ and beta rhythms, improving input signal quality for any downstream model.

4. Batch Normalization for Faster Convergence: Insert batch normalization after convolutional layers to stabilize training, reduce overfitting, and accelerate optimization, making the system more efficient for real-time or resource-constrained deployments.

5. Reduced Model Complexity with Higher Accuracy: Replace deeper recurrent networks with a single bi-LSTM layer after CNN features, achieving better accuracy than increasing LSTM layers/parameters—this yields a lighter, faster model suitable for embedded BCI hardware.

What the Improved AI System Can Do:

• Decode two- and three-class motor imagery tasks (e.g., left/right hand, foot) with 85–92% accuracy across private and public datasets, outperforming LDA, SVM, CNN-only, and RNN-only baselines.

• Operate without subject-specific calibration, enabling immediate use for new users (subject-independent decoding).

• Process raw EEG signals directly, eliminating the need for manual feature selection, thus reducing expert intervention and deployment time.

• Maintain high performance under noisy, low-SNR conditions (e.g., real-world clinical or home rehabilitation settings) due to wavelet denoising and frequency-specific filtering.

• Support real-time or near-real-time BCI applications (e.g., neurorehabilitation, assistive devices, gaming) because of its efficient architecture and batch-normalized training.

Motor imagery (MI) brain--computer interfaces (BCIs) have emerged as a promising approach for establishing flexible communication pathways between the human brain and external devices, particularly for individuals affected by stroke or neurodegenerative disorders. Reliable decoding of motor-imagery electroencephalography (MI-EEG) remains challenging because EEG recordings contain substantial noise and exhibit complex, weakly informative relationships with the underlying brain activity. Although deep learning provides an effective means of learning representations directly from EEG signals, its application to MI-EEG feature learning remains comparatively limited. This study introduces a hybrid deep-learning architecture that integrates a convolutional neural network (CNN) with a bidirectional long short-term memory (bi-LSTM) network. The CNN is used to learn high-level spatial and temporal representations directly from raw MI-EEG recordings, whereas the bi-LSTM models temporal dependencies and relationships among the extracted features. The proposed approach is evaluated using both a publicly available dataset and a privately acquired dataset obtained with an EEG acquisition system. The experimental results indicate that the CNN&bi-LSTM architecture provides robust performance for both two- and three-class motor-imagery classification and demonstrates promising subject-independent decoding capability across the evaluated methods.