Event-Related Spectrogram Representation of EEG for CNN-Based P300 Speller

P300 speller is one of the most popular applications for electroencephalography (EEG) features extraction and classification. It is used for enabling paralyzed people to communicate with the outer world by inputting text using EEG signals, evoked by their brain. Classical P300 speller uses EEG time-series classification, which does not cover all of the frequency bands of brain activity and is very subject-dependent. In order to make the system more stable and robust, frequency domain spectrograms such as intertrial coherence (IT C) and event- related spectral perturbation (ERSP) are generated from EEG data of amyotrophic lateral sclerosis (ALS) patients. The obtained spectrograms are further classified using a convolutional neural network (CNN). CNN-based ensemble model is proposed for classifying both ITC and ERSP in order to achieve more trusted results. Classification of EEG data in the form of images can make the whole process much more uniform, robust and independent from the data acquisition properties.