Motor Imagery BCI Classification

Overview

Motor imagery (MI) BCI is a paradigm where a person imagines performing a movement — without actually moving — and the resulting neural activity in the motor cortex is decoded from EEG signals to produce a control command. MI BCIs are used in rehabilitation robotics, assistive technology for paralysed patients, and neuroscience research. The central challenge is that EEG is low-SNR, non-stationary, and highly individual-specific: patterns that decode well for one subject require re-calibration for another.

This project benchmarks three classification approaches on BCI Competition 2005 Dataset I (a standard, well-characterised benchmark): a classical signal processing pipeline (CSP+SVM), a compact end-to-end deep learning architecture (EEGNet), and a custom 1D CNN. All three use 5-fold stratified cross-validation for fair comparison and to detect overfitting.

Dataset

Preprocessing Pipeline

Classifier 1: Common Spatial Patterns + SVM

Common Spatial Patterns (CSP) is a classical spatial filter algorithm that maximises the variance ratio between two classes in EEG. Given the covariance matrices Σ₁ (class 1) and Σ₂ (class 2), CSP finds the filter matrix W such that:

The first and last columns of W correspond to the directions of maximum variance for class 1 and class 2 respectively. Log-variance of the filtered signals (log(var(W^T·x))) forms the feature vector fed to an SVM with RBF kernel. CSP+SVM achieved 81.2% accuracy.

Classifier 2: EEGNet

EEGNet (Lawhern et al. 2018) is a compact CNN designed specifically for EEG classification. Its architecture:

• Block 1: Temporal convolution (1×64, captures spectral content) → Depthwise spatial convolution (C×1 per channel pair, spatial filtering) → BatchNorm → ELU → AvgPool → Dropout
• Block 2: Separable convolution (1×16, learns temporal dynamics of spatial features) → BatchNorm → ELU → AvgPool → Dropout
• Classifier: Flatten → Dense → Softmax

EEGNet has ~2,500 parameters (very compact), which is important for EEG where limited training data makes large models prone to overfitting. Achieved 83.4% accuracy.

Classifier 3: Custom 1D CNN

The custom 1D CNN treats each channel's time series independently through depthwise convolutions, then fuses across channels. Architecture:

• Input: (64 channels × 750 time points)
• 4× temporal Conv1D blocks (increasing kernel sizes: 32, 16, 8, 4 samples)
• Batch normalisation + ReLU + MaxPool after each block
• Channel attention module (squeeze-excitation) for spatial feature weighting
• Global average pooling → FC(128) → Dropout(0.5) → FC(2) → Softmax

Achieved 83.8% accuracy — the best of the three, marginalyl ahead of EEGNet. The channel attention mechanism proved important: ablation without it reduced accuracy by 2.1%.

Results

Fig. 2. Raw EEG data overview: 64-channel recording showing characteristic μ-band (8–12 Hz) and β-band (13–30 Hz) oscillations in motor cortex channels.

Fig. 3. Time-frequency spectrograms for representative channels (C3, Cz, C4). Event-related desynchronisation (ERD) visible as power decrease during motor imagery.

Fig. 4. Accuracy comparison across the three classifiers. Error bars show std across the 5 cross-validation folds. 1D CNN: 83.8%, EEGNet: 83.4%, CSP+SVM: 81.2%.

Fig. 5. Confusion matrices for all three classifiers. All show stronger performance on right-hand imagery — consistent with the literature on this dataset.

Fig. 6. Training and validation accuracy curves for the 1D CNN. Convergence within 80 epochs; no visible overfitting with dropout regularisation.

Fig. 7. CSP spatial filter patterns: topographic maps of the first and last filter columns. Clear lateralisation to left/right motor cortex (C3/C4 areas).