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

The brain-computer interface (BCI) establishes a non-muscle channel that enables direct communication between the human body and an external device. Electroencephalography (EEG) is a popular non-invasive technique for recording brain signals. It is critical to process and comprehend the hidden patterns linked to a specific cognitive or motor task, for instance, measured through the motor imagery brain-computer interface (MI-BCI). A significant challenge is presented by classifying motor imagery-based electroencephalogram (MI-EEG) tasks, given that EEG signals exhibit nonstationarity, time-variance, and individual diversity. Obtaining good classification accuracy is also very difficult due to the growing number of classes and the natural variability among individuals. To overcome these issues, this paper proposes a novel method for classifying EEG motor imagery signals that extracts features efficiently with Minimally Random Convolutional Kernel Transform (MiniRocket), a linear classifier then uses the extracted features for activity recognition. Furthermore, a novel deep learning based on Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) architecture to serve as a baseline was proposed and demonstrated that classification via MiniRocket’s features achieves higher performance than the best deep learning models at lower computational cost. The PhysioNet dataset was used to evaluate the performance of the proposed approaches. The proposed models achieved mean accuracy values of 98.63% and 98.06% for the MiniRocket and CNN-LSTM, respectively. The findings demonstrate that the proposed approach can significantly enhance motor imagery EEG accuracy and provide new insights into the feature extraction and classification of MI-EEG.

💡 Research Summary

The paper addresses the challenging problem of classifying motor‑imagery (MI) electroencephalogram (EEG) signals for brain‑computer interface (BCI) applications. It proposes and empirically compares two fundamentally different pipelines: (1) a feature‑extraction‑then‑linear‑classification approach based on MiniRocket, a recent time‑series transformation that uses a large set of minimally random 1‑D convolutional kernels, followed by a ridge classifier; and (2) a conventional deep‑learning baseline that combines a convolutional neural network (CNN) with a long short‑term memory (LSTM) recurrent network.

Data are drawn from the PhysioNet MI‑EEG dataset recorded with the BCI2000 system. The dataset contains recordings from 109 subjects, each performing four MI tasks (left‑hand, right‑hand, both‑hands, both‑feet) with 21 trials per task, sampled at 160 Hz. Prior to any modeling, the authors apply independent component analysis (ICA) to remove ocular and muscular artifacts and isolate the μ (8‑14 Hz) and β (14‑30 Hz) bands, which are known to carry the most discriminative information for MI.

MiniRocket works by generating a fixed pool of random convolutional kernels (10 000 in the experiments) that are applied to each EEG trial (640 samples per channel). After each convolution, simple non‑linear pooling operations (e.g., proportion of positive values, max‑pool) produce a compact 9‑dimensional feature vector per trial. Because the kernels are not learned, the transformation is extremely fast and memory‑efficient. The resulting feature matrix is fed to a ridge regression classifier, which learns a linear decision boundary for the four MI classes.

The CNN‑LSTM baseline first passes the raw EEG through two 1‑D convolutional layers (16 and 32 filters, kernel size 3) with ReLU activation and max‑pooling, followed by dropout layers (rates 0.25–0.5) to mitigate over‑fitting. The extracted feature maps are then processed by three stacked LSTM layers, each with 100 hidden units, to capture long‑range temporal dependencies. Finally, fully‑connected dense layers with softmax output produce class probabilities.

Experimental results show that MiniRocket + ridge achieves a mean classification accuracy of 98.63 % (standard deviation ≈ 0.42 %) across ten subject‑wise splits, while the CNN‑LSTM reaches 98.06 %. In terms of computational cost, MiniRocket requires less than one second for training on the entire dataset and only a few milliseconds for inference per trial, whereas the CNN‑LSTM needs several hours of GPU training and roughly 150 ms per inference. Memory consumption follows the same trend, with MiniRocket using a fraction of the GPU memory needed by the deep network.

The authors draw several key insights: (i) the random‑kernel approach of MiniRocket captures the essential spatio‑temporal patterns of MI‑EEG without any learned convolutional filters, demonstrating that sophisticated deep architectures are not mandatory for high accuracy when a large, diverse kernel bank is used; (ii) a simple linear ridge classifier is sufficient to separate the four MI classes once the MiniRocket features are available, highlighting the power of high‑dimensional random projections; (iii) CNN‑LSTM remains valuable for scenarios where end‑to‑end learning of spatial and temporal representations is desired, but it suffers from higher computational demand and a greater risk of over‑fitting, especially with limited training data; (iv) ICA‑based preprocessing and explicit μ/β band extraction improve signal‑to‑noise ratio and benefit both pipelines.

The paper concludes that MiniRocket‑ridge constitutes a highly efficient, accurate, and easily deployable solution for real‑time MI‑BCI systems, particularly in resource‑constrained environments such as wearable or embedded devices. Future work is suggested in three directions: extending the approach to additional frequency bands (e.g., γ), evaluating performance with varying numbers of EEG channels, and integrating online adaptation mechanisms to personalize the classifier for individual users.