Comparison of Deep Learning and Traditional Machine Learning Classification Performance in a Steady State Visual Evoked Potential Based Brain Computer Interface

Abstract

Brain-computer interfaces (BCIs) offer a very high potential to help those who cannot use their organs properly. In the literature, many electroencephalogram based BCIs exist. Steady state visual evoked potential (SSVEP) based BCIs provide relatively higher accuracy values which make them very popular in BCI research. Recently, deep learning (DL) based methods have been used in electroencephalogram classification problems and they had superior performance over traditional machine learning (ML) methods, which require feature extraction step. This study aimed at comparing the performance of DL and traditional ML based classification performance in terms of stimuli duration, number of channels, and number of trials in an SSVEP based BCI experiment. In the traditional approach canonical correlation analysis method was used for the feature extraction and then three well-known classifiers were used for classification. In DL-based classification, spatio-spectral decomposition (SSD) method was integrated as a preprocessing step to extract oscillatory signals in the frequency band of interest with a convolutional neural network structure. Obtained offline classification results show that proposed DL approach could generate better accuracy values than traditional ML-based methods for short time segments (< 1 s). Besides, use of SSD as a preprocessing step increased the accuracy of DL classification. Superior performance of proposed SSD based DL approach over the traditional ML methods in short trials shows the feasibility of this approach in future BCI designs. Similar approach can be used in other fields where there are oscillatory activity in the recorded signals.