Time-Frequency Representations for EEG Artifact Classification with CNNs

Electroencephalogram (EEG) is used in the diagnosis of a wide range of neurological conditions and in cognitive science research. However, one important problem while using EEG is the presence of different types of artifacts. Some of the well-known types of artifacts are caused due to chewing, electrode pops, eye movements, muscle movements, and shivering. These artifacts can lead to misinterpretation of brain signals and hamper the precise encoding of brain activities. To develop methods to mitigate the artifacts, it is useful to identify the artifacts, as the characteristics of specific artifacts can help in devising targeted methods. However, manual identification of EEG artifacts is time-consuming and tedious, and automated classifiers can be very useful to improve the identification efficiency. In this work, we propose three different deep learning based methods to recognize the artifacts, and to discriminate them from the normal EEG signals. Two out of the three methods are based on frequency domain representation of EEG signals, while one works with raw EEG signals. The frequency domain representations include Fourier transform and spectrogram. Importantly, each of the three representations is associated with the appropriately structured CNN. We provide some insights about why some rep-resentations may be more useful and, experimentally demonstrate improvements over some recent methods on a standard large dataset containing a variety of artifacts.