Investigation on CNN-Based State Classification Towards BCI Application Using Amplitude Probability Density Distribution

Various kinds of analysis including fast Fourier transform (FFT) were widely used for the classification of electroencephalogram (EEG) based human interfaces. However, the morphological characteristics of EEG waveform were rarely used, since the EEG waveform is thought to show no significant meaning due to its stochastic features. The authors have studied on SSVEP-based BCI for disabled patients. The objective of this study is to verify feasibility of amplitude probability density distribution (APD) used as a feature contributing classification EEG. In this study, the amplitude probability density distribution, which indicated as the index of morphological characteristics of EEG, was applied for state classification using deep learning. CNN was introduced to construct the model of deep learning, classify the obtained data calculated by introduced novel APD method and FFT in order to compare the feasibility. The data were obtained from EEG recorded when subjects were presented flashing light with low stimulus luminosity reversed at 20 and 60 Hz. EEG measurement was conducted in shield room and 9 healthy adulthood male subjects participated in this study. As a result, the case of EEG spectrum by FFT as the control data, the classification accuracy was 85.81%, while using APD yielded 87.98%. The classification accuracy in both analyses showed almost similar result. To conclude, apart from the traditional frequency characteristic, it is feasible in utilization of morphological information in classifying EEG characteristics obtained from two different frequencies. Some problems of implementation for the BCI and efficacy of present method will be discussed.