⚡ electrical engineering

Motor Imagery EEG Signal Classification Using Minimally Random Convolutional Kernel Transform and Hybrid Deep Learning

This paper proposes a novel motor imagery EEG classification framework that leverages the Minimally Random Convolutional Kernel Transform (MiniRocket) for efficient feature extraction and demonstrates that it outperforms a hybrid CNN-LSTM deep learning baseline in both accuracy (98.63%) and computational efficiency on the PhysioNet dataset.

Original authors: Jamal Hwaidi, Mohamed Chahine Ghanem

Published 2026-02-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jamal Hwaidi, Mohamed Chahine Ghanem

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 massive, bustling radio station. Every time you think about moving your hand or foot, even if you don't actually move it, your brain sends out a specific "broadcast" signal. This is called Motor Imagery (MI).

The goal of this research is to build a super-smart translator that can listen to these brain broadcasts and figure out exactly what you are imagining doing. This technology is the backbone of Brain-Computer Interfaces (BCI), which could one day let paralyzed people control robotic arms or wheelchairs just by thinking.

However, there's a problem: Brain signals are messy. They are like a radio station playing in a crowded room with static, interference, and different voices all talking at once. Making sense of them is incredibly hard.

This paper presents two new "translators" to solve this puzzle. Let's break them down using simple analogies.

The Two Translators

The researchers tested two different approaches to decode these brain signals:

1. The "Super-Fast Detective" (MiniRocket)

Think of the brain signal as a long, messy string of beads. Traditional methods try to untangle every single knot to find a pattern, which takes forever.

MiniRocket is like a detective who doesn't try to untangle the whole string. Instead, they use a special, pre-made "template" (a random kernel) to quickly scan the string.

How it works: It throws thousands of these templates at the data. It doesn't learn them; it just uses them to instantly spot specific shapes or patterns (like "Oh, this part looks like a left-hand movement!").
The Magic: It's incredibly fast and efficient. It's like using a metal detector on a beach instead of digging every inch of sand with a spoon. It finds the "treasure" (the important features) without doing all the heavy lifting.
The Result: This method was the winner. It achieved 98.63% accuracy and did it much faster than the other method.

2. The "Deep-Dive Student" (CNN-LSTM Hybrid)

This approach is like a very dedicated student who wants to understand the entire story, not just the clues.

The CNN (Convolutional Neural Network): This part is like a microscope. It looks at the brain signals to find spatial patterns (where the signal is coming from on the scalp). It asks, "Is the signal stronger on the left side or the right side?"
The LSTM (Long Short-Term Memory): This part is like a time-traveling historian. It looks at the temporal patterns (how the signal changes over time). It asks, "Did the signal start low and then spike?"
The Hybrid: By combining the microscope and the historian, this model tries to learn everything about the signal from scratch.
The Result: It did very well too (98.06% accuracy), but it required much more computer power and time to "study" the data compared to the Super-Fast Detective.

The Experiment: The "Brain Gym"

To test these translators, the researchers used a public dataset called PhysioNet. Imagine this as a giant gym where 10 different people (subjects) went through a workout routine.

The Workout: Each person sat in a chair and imagined four different things:
1. Moving their Left Fist.
2. Moving their Right Fist.
3. Moving Both Fists.
4. Moving Both Feet.
The Challenge: The computer had to listen to the brain waves and guess which of the four exercises the person was imagining.

The Big Takeaway

The study found that you don't always need the most complex, heavy machinery to get the best results.

The "Deep-Dive Student" (CNN-LSTM) is powerful and impressive, but it's like driving a Ferrari to go to the grocery store. It works, but it's expensive and slow.
The "Super-Fast Detective" (MiniRocket) is like a sleek, electric scooter. It gets you to the destination faster, uses less energy, and actually got a slightly better score (98.63% vs. 98.06%).

Why This Matters

This research is a big step forward for Brain-Computer Interfaces because:

Speed: It proves we can decode brain signals very quickly, which is crucial for real-time control (like moving a robotic arm right now).
Efficiency: It shows we don't need super-computers to make these systems work. This makes the technology cheaper and more accessible for hospitals and homes.
Accuracy: Getting to nearly 99% accuracy means these systems are becoming reliable enough for real-world medical use.

In short, the authors found a way to listen to the brain's "radio station" clearly and quickly, proving that sometimes the simplest, smartest tool is better than the most complicated one.

1. Problem Statement

The paper addresses the challenge of classifying Motor Imagery (MI) Electroencephalography (EEG) signals for Brain-Computer Interface (BCI) applications. Key challenges include:

Signal Characteristics: EEG signals are non-stationary, time-variant, and exhibit high inter-subject diversity.
Low Signal-to-Noise Ratio (SNR): The signals are microvolt-level and prone to artifacts.
Data Scarcity: Deep learning models often require large datasets, which are difficult to obtain for specific EEG tasks.
Computational Cost: State-of-the-art (SOTA) time-series classification methods often suffer from high computational complexity, making them difficult to train on smaller datasets or impractical for real-time applications.
Feature Extraction: Traditional methods often rely on manual feature engineering (e.g., CSP, PSD) which may miss dynamic temporal patterns or fail to generalize across subjects.

2. Methodology

The authors propose a comparative framework utilizing two distinct approaches on the PhysioNet EEG Motor Imagery Dataset (109 subjects, 4 MI tasks: left fist, right fist, both fists, both feet).

A. Data Preprocessing

Source: PhysioNet dataset (160 Hz sampling rate).
Segmentation: Trials were segmented into 4-second windows (640 samples per electrode).
Frequency Filtering: Independent Component Analysis (ICA) was used to isolate $\mu$ (8–14 Hz) and $\beta$ (14–30 Hz) bands, which are most discriminative for MI.
Normalization: Data was averaged using a common reference, and 3-second pre-trial intervals were removed.

B. Approach 1: MiniRocket with Linear Classifier

Core Concept: Utilizes Minimally Random Convolutional Kernel Transform (MiniRocket), a highly efficient time-series classifier.
Mechanism:
- Applies 10,000 random convolutional kernels (with fixed weights, biases, and dilations) to the input signal.
- Extracts features using the Proportion of Positive Values (PPV) and max pooling.
- Unlike standard CNNs, the kernel weights are not learned; they are fixed and random, drastically reducing training time.
- The transformed features are fed into a Ridge Linear Classifier.
Advantage: Significantly lower computational complexity ( $O(I \cdot K \cdot S)$ ) compared to deep learning, while retaining high accuracy.

C. Approach 2: Hybrid CNN-LSTM Deep Learning Model

Architecture: A 13-layer hybrid network designed to capture both spatial and temporal dependencies.
- Spatial Feature Extraction: Two 1D Convolutional layers (16 and 32 filters) with ReLU activation to learn local patterns.
- Regularization: Dropout layers (0.5 after conv, 0.25 after dense) and L2 regularization to prevent overfitting.
- Temporal Feature Extraction: A Long Short-Term Memory (LSTM) layer with 100 units to model sequential dependencies and long-term memory.
- Classification: Fully connected (Dense) layers with ReLU and a final Sigmoid activation for 4-class output.
Training: Optimized using the Adam algorithm with a learning rate of $1 \times 10^{-5}$ and batch size of 64 (training) / 400 (validation plots).

3. Key Contributions

Novel Feature Extraction: Application of MiniRocket to MI-EEG classification, demonstrating that random convolutional kernels can outperform complex deep learning models in terms of efficiency and accuracy.
Hybrid Baseline: Development of a robust CNN-LSTM hybrid architecture specifically tuned for MI-EEG, serving as a high-performance baseline for comparison.
Frequency-Specific Preprocessing: Strategic use of ICA to isolate $\mu$ and $\beta$ bands before feeding data into the classifiers, enhancing signal relevance.
Comprehensive Evaluation: A rigorous comparison on the PhysioNet dataset involving 10 subjects, 10-fold cross-validation, and analysis of confusion matrices, ROC curves, and training dynamics.

4. Experimental Results

The models were evaluated on the PhysioNet dataset (10 subjects, 4 classes).

Accuracy Performance:
- MiniRocket: Achieved a mean accuracy of 98.63%.
- CNN-LSTM: Achieved a mean accuracy of 98.06%.
- Observation: MiniRocket slightly outperformed the deep learning model while requiring significantly less computational resources.
Subject Performance:
- Best subject (S8) achieved 99.44% (MiniRocket) and 98.56% (CNN-LSTM) for the left fist task.
- The lowest performing subject (S3 for MiniRocket, S10 for CNN-LSTM) still maintained high accuracy (>90% across most tasks).
Comparison with State-of-the-Art:
- The proposed MiniRocket approach (98.63%) significantly outperformed previous studies using the same dataset, which ranged from 74.90% (SVM) to 97.71% (DSCNN+ELM).
Robustness: Both models showed stable convergence with low loss and high precision/recall across all four MI classes (Left Fist, Right Fist, Both Fists, Both Feet).

5. Significance and Conclusion

Efficiency vs. Accuracy: The study demonstrates that MiniRocket offers a superior trade-off for MI-BCI systems. It achieves SOTA accuracy with a fraction of the computational cost and training time required by deep learning models.
Practical Application: The high accuracy and low latency of the MiniRocket approach make it highly suitable for real-time BCI applications where computational power is limited (e.g., portable devices).
Future Directions: The authors suggest that while the offline results are promising, future work will focus on online experimentation, adapting the approach to different experimental paradigms, and designing specific data collection protocols to further validate the model's generalizability.

In summary, this paper establishes that randomized kernel transforms (MiniRocket) are a highly effective, computationally efficient alternative to complex deep learning architectures for classifying motor imagery EEG signals, achieving record-breaking accuracy on standard benchmarks.