Motor imagery EEG signal classification using minimally random convolutional kernel transform and hybrid deep learning

Comprehensive Summary

This study investigates improved methods for classifying motor-imagery (MI) electroencephalography (EEG) signals, which are widely used in brain–computer interface (BCI) systems. Motor imagery refers to the mental simulation of movements (e.g., imagining moving the left or right hand), which produces measurable neural activity patterns. Accurate classification of these signals is essential for translating brain activity into commands for assistive technologies. The researchers evaluated a computational framework combining MiniRocket feature extraction with deep learning architectures, including convolutional neural networks (CNN) and long short-term memory (LSTM) networks. MiniRocket is a time-series feature-extraction algorithm that applies thousands of random convolutional kernels to EEG signals and computes summary features (such as the proportion of positive values) to efficiently represent temporal patterns. Compared with earlier algorithms like ROCKET, MiniRocket maintains a high-dimensional feature representation while reducing computational cost. The study used publicly available EEG motor-imagery datasets, particularly the PhysioNet MI-EEG dataset, with preprocessing steps including band selection (μ and β rhythms), signal segmentation, and independent component analysis to isolate relevant neural signals. Data were split into training, validation, and testing sets while preventing subject-level data leakage. Results showed that combining MiniRocket with hybrid deep-learning models (CNN–LSTM) achieved strong classification performance for motor-imagery EEG signals while maintaining computational efficiency. The model captured both spatial features (via CNN layers) and temporal dependencies (via LSTM or dilated convolutions). The authors also observed substantial inter-subject variability, meaning classification accuracy differed across individuals, highlighting a major challenge in BCI systems. Overall, the study demonstrates that efficient time-series feature extraction combined with deep neural networks can significantly improve motor-imagery EEG decoding, potentially enabling faster and more scalable brain-computer interface systems.

Outcomes and Implications

This research has important implications for the development of clinical brain–computer interface technologies, particularly for individuals with severe motor impairments. Accurate decoding of motor-imagery EEG signals could allow patients with conditions such as spinal cord injury, stroke, or neurodegenerative diseases to control assistive devices using only brain activity. Improved classification algorithms like MiniRocket-based models can reduce computational cost while maintaining high accuracy, which is critical for real-time BCI applications such as robotic prosthetic control, wheelchair navigation, or communication interfaces for patients with paralysis (e.g., locked-in syndrome). Faster algorithms also make it more feasible to deploy BCI systems in portable or wearable clinical devices. Additionally, the study highlights the challenge of inter-subject variability, meaning that neural signals differ substantially between individuals. Addressing this variability is essential for creating clinically viable BCI systems that can adapt to different patients without extensive retraining. Future work integrating multimodal neural recordings (e.g., EEG with other physiological signals) may improve robustness and reliability for clinical deployment.