Time-Domain Versus Frequency-Embedded EEG Sequences for Sensorimotor BCI Using 1D-CNN

This study proposed a motor imagery (MI) classification pipeline featuring a 1−dimensional convolutional neural network (1D-CNN) with different time/frequency feature representation techniques. The objective was to classify right hand (RH) versus right foot (RF) MI tasks in both intra- and inter-subject (pairwise and pooled) BCI settings using a 1D-CNN architecture trained on time-domain bandpass filtered electroencephalography (EEG) signals, frequency-embedded power spectral density (PSD) and cross-power spectral density (CPSD) sequences. The EEG signals were bandpass filtered with 4 Hz and 32 Hz cut-off frequencies, and PSD/CPSD sequences were estimated in the same frequency range. Thus, the number of input channels for 1D-CNN was N, N or $N\times N$ for EEG signals, PSD or CPSD sequences. We used dataset IVa from BCI Competition III in 5−fold cross-validation settings to evaluate intra-subject, inter-subject (pairwise) and inter-subject (pooled) BCI classification accuracies. We compared the performance of the proposed methods with classification algorithms featuring common spatial patterns (CSP) for benchmarking. The best overall classification accuracies (%) for intra-subject, inter-subject (pairwise) and inter-subject (pooled) BCIs were $86.57\pm 11.69$ , $70.80\pm 9.21$ and $76.61\pm 12.37$ using 1D−CNN with time-domain EEG signals. The average classification accuracies using 1D-CNN with frequency-embedded PSD sequences were $82.57\pm 10.20$ , $69.32\pm 7.46$ and $71.32\pm 7.96$ for intra-subject, inter-subject (pairwise) and inter-subject (pooled) BCIs. The proposed time/frequency feature representation techniques with 1D-CNN outperformed CSP-based algorithms (p-value <0.05). The comparative results suggest the utility of the proposed methods for MI classification, especially for a fully zero-training inter-subject BCI.