🤖 machine learning

EEG Decoding Using CNN and LSTM Network

This paper proposes a hybrid deep-learning architecture combining a Convolutional Neural Network (CNN) and a bidirectional Long Short-Term Memory (bi-LSTM) network to robustly decode motor imagery EEG signals for both subject-dependent and subject-independent classification tasks.

Original authors: Athanasios Karagounis

Published 2026-08-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Athanasios Karagounis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a bustling city where billions of tiny messengers (neurons) are constantly shouting messages to one another. Sometimes, these messages are about moving your hand, or maybe just imagining moving your hand. A Brain-Computer Interface (BCI) is like a super-advanced walkie-talkie that tries to listen to these whispers and turn them into commands for a computer, a robot, or a video game. This is a huge deal for people who can't move their bodies due to illness or injury, giving them a new way to talk to the world.

But there's a catch: the brain's signals are incredibly messy. It's like trying to hear a single person whisper in a stadium full of cheering fans. The signals are weak, noisy, and hard to understand. For a long time, scientists had to manually clean up this noise and guess which parts of the signal mattered, kind of like trying to find a needle in a haystack by hand. Recently, computers have started using "deep learning"—a type of artificial intelligence that learns patterns on its own—to do this job better. However, even these smart computers sometimes struggle to catch the full story because brain signals change over time, and the computer needs to understand both where the signal comes from and when it happens.

This is where the study by Athanasios Karagounis steps in. The researcher built a new kind of "brain-reading" machine that combines two powerful AI tools: a Convolutional Neural Network (CNN) and a Bidirectional Long Short-Term Memory (bi-LSTM) network. Think of the CNN as a super-sleuth that scans the brain's electrical map to find the most important spots, filtering out the static noise just like a radio tuner finding a clear station. Once the CNN finds the good clues, the bi-LSTM acts like a time-traveling detective. It doesn't just look at the clues in the order they happened; it looks at the past and the future of the signal at the same time to understand the full story of what the person is imagining.

The paper tests this new hybrid system on two different groups of data: a private set of recordings from six healthy people and several public datasets from other researchers. The goal was to see if the computer could tell the difference between people imagining moving their left hand, their right hand, or even their tongue. The results were promising. The new CNN&bi-LSTM team beat out older, standard methods like Linear Discriminant Analysis (LDA) and Support Vector Machines (SVM). In tests with the private data, the new system got about 85.4% of the answers right on average, while the older methods hovered around 69% to 79%. Even more interestingly, the system worked well even when the amount of training data was cut down, suggesting it's robust and doesn't need a massive library of examples to learn.

The researchers found that the system learned best when it focused on 40 specific electrodes (sensors) on the head, rather than using all 64 available, which actually helped avoid "overthinking" the data. They also discovered that the AI didn't just memorize the signals; it started to recognize complex patterns that looked like waves, getting more sophisticated as it dug deeper into the layers of its own "brain." While the improvement over existing methods was steady rather than a massive explosion, the study suggests that combining spatial filtering (finding the right spot) with bidirectional time modeling (understanding the flow of time) is a smarter way to decode our thoughts. The paper concludes that this approach offers a solid, unified way to turn messy brain waves into clear commands, potentially paving the way for better rehabilitation tools for people who need them most.

Technical Summary: EEG Decoding Using CNN and LSTM Network

Problem Statement
Motor imagery (MI) brain–computer interfaces (BCIs) offer a promising communication pathway for individuals with stroke or neurodegenerative disorders by translating imagined movements into control signals without requiring overt physical motion. However, reliable decoding of motor-imagery electroencephalography (MI-EEG) remains challenging due to the substantial noise inherent in EEG recordings and the complex, weakly informative relationships between the signals and underlying brain activity. While deep learning offers a means to learn representations directly from raw signals, its application to MI-EEG feature learning has been comparatively limited. Furthermore, conventional deep-learning pipelines relying solely on individual Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) often require significant manual effort for feature engineering, parameter optimization, and model training, failing to fully capture the necessary spatial and temporal dependencies.

Methodology
The study proposes a hybrid deep-learning architecture that integrates a Convolutional Neural Network (CNN) with a Bidirectional Long Short-Term Memory (bi-LSTM) network to address these limitations.

Data Acquisition and Preprocessing: The framework was evaluated using a private dataset (D1) acquired with 64 electrodes and three public datasets (D2–D4) with varying electrode counts (32 or 64). Preprocessing involved referencing signals to the vertex electrode (Cz), selecting principal MI-EEG electrode locations (e.g., C3, C4, P3, P4), applying wavelet-based denoising, and band-pass filtering within the 8–23 Hz range to isolate $\mu$ and $\beta$ rhythms.
CNN Feature Extraction: The CNN component serves as a spatial filter designed to learn high-level spatial and temporal representations directly from raw MI-EEG recordings. It employs eight spatial filters with a kernel size of $[36 \times 1]$ to transform original EEG channels into task-oriented feature representations. This layer performs learned spatial filtering analogous to band-pass filtering and energy extraction, reducing reliance on manual feature selection.
Batch Normalization: To improve generalization and accelerate optimization, batch normalization (BN) is applied after feature extraction. This normalizes inputs to a mean of zero and a standard deviation of one, mitigating internal covariate shift.
Bidirectional LSTM Modeling: The extracted features are processed by a bi-LSTM network to model temporal dynamics and dependencies. Unlike standard LSTMs that only utilize preceding inputs, the bi-LSTM processes the sequence in both forward and backward directions. This allows the network to incorporate contextual information from both earlier and later observations, capturing complementary dependencies across different timescales. The network utilizes standard LSTM gates (input, forget, output) and cell states to mitigate vanishing gradient problems.
Training Strategy: The model is trained using cross-validation with Mean Squared Error (MSE) as the objective function, chosen over cross-correlation due to the presence of resting-state intervals in the data. Supervised pre-training is employed to initialize weights using LASSO-derived coefficients, providing a solution related to least-squares regression before optimizing for temporal correlations.

Key Contributions

Hybrid Architecture: The paper introduces a CNN&bi-LSTM architecture specifically designed to retain informative long-range temporal patterns while improving the classification performance of MI-EEG-based BCIs.
End-to-End Learning: The framework learns spatial and temporal dependencies directly from raw EEG measurements. The CNN extracts features without manual engineering, while the bi-LSTM models sequential dynamics to improve the representation of temporal dependencies.
Comprehensive Evaluation: The approach is extensively assessed using both a private dataset and multiple publicly available datasets, demonstrating performance across different subject groups and experimental configurations.

Experimental Results
The proposed method was compared against Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), standard CNNs, and RNNs.

Private Dataset (D1): The CNN&bi-LSTM model achieved a mean classification accuracy of 85.4% across seven subjects. This outperformed LDA (69.5%), SVM (73.0%), CNN (75.9%), and RNN (79.5%). The highest subject-specific accuracy reached 92.5%.
Public Datasets (D2–D4): The hybrid model consistently outperformed alternative machine learning and recurrent architectures. On D2, D3, and D4, the proposed model achieved accuracies of 81.80%, 87.97%, and 85.32%, respectively. These results were superior to single-layer LSTMs, multi-layer RNNs, and ST-CNNs, indicating that combining convolutional feature extraction with bidirectional temporal modeling is more effective than simply increasing network depth.
Robustness: Experiments indicated that the architecture is relatively robust to reductions in training data size, maintaining approximately 89% accuracy with only 60% of available data. Additionally, performance peaked with 40 electrodes, suggesting that excessive channel counts may lead to overfitting.

Significance and Claims
The paper claims that the proposed 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 results demonstrate that the method can effectively emphasize informative components of MI-EEG signals and achieve competitive classification performance across diverse datasets and subjects. The author posits that the use of bidirectional recurrent modeling represents a potential alternative to conventional spectral-feature-based EEG analysis. They suggest this approach may support future research and practical applications in MI-EEG-based rehabilitation systems, offering a robust solution for subject-independent decoding.

More like this