User-Independent SSVEP BCI Using Complex FFT Features and CNN Classification

User-independent Brain Computer Interfaces (BCIs) have gained increased attention in recent years for their attractive feature of having minimal or zero calibration. BCIs based on steady state visual evoked potentials (SSVEP) have the most favorable characteristics for developing user-independent BCIs. In this study, we proposed the use of complex Fast Fourier Transform (FFT) features as input to a Convolutional Neural Network (CNN) for classifying SSVEP responses without user specific training. Our proposed method (C-CNN) was tested on two SSVEP datasets, a 7-class SSVEP dataset with 14 participants and a publicly available 12-class SSVEP dataset with 10 participants. We compared the proposed method with Canonical Correlation Analysis (CCA) and CNN classification using the magnitude spectrum features (M-CNN) of SSVEP. Results showed that the proposed C-CNN outperforms CCA and M-CNN, with an accuracy significantly higher than CCA, in both 7-class and 12-class SSVEP datasets. Overall accuracies comparing C-CNN vs M-CNN vs CCA were 79.42% vs. 69.60% vs. 67.91 (7-class) and 81.6% vs. 70.60% vs. 62.7% (12-class) with a data length of 1 second. The results suggest that by using the complex FFT features, the CNN learns to use both frequency and phase related information to classify SSVEP responses. With 1 second data length, user-independent training and a simple model, the proposed method is suitable for developing practical BCI systems that can enhance human-computer interactions.