The ear-EEG artifact as a predictor of motion states on a treadmill: A bispectral analysis approach

In this paper the problem of human motion states classification is studied, namely standing, walking and running, using machine learning techniques applied to ear-electroencephalogram (ear-EEG) data. Ear-EEG sensors are a promising alternative to traditional EEG devices as they are lightweight, easy to use and less intrusive. By exploiting the advantages of ear-EEG, this work aims to investigate its effectiveness in motion classification tasks utilizing the movement artifact component. Bispectral analysis was applied on an existing ear-EEG dataset, acquired during a mobile brain computer interface (BCI) experimental workflow. Through this approach, the demonstration of the bispectrum-derived features efficacy in classifying motion states was aimed. Experiments were conducted to evaluate the performance of different machine learning algorithms using nested K-fold cross validation. In addition, model evaluation and hyperparameter tuning were performed using nested Leave-One-Subject-Out (LOSO) cross validation. The experiments were repeated twice, using a different re-reference method for the data in each iteration. A performance of 82.72% in test accuracy was achieved with Random Forest and the all-mean re-reference method. The outcome of this study highlights the usability of ear-EEG artifacts and provides valuable insights for future application of ear-EEG and bispectral analysis in the field of human motion, with clinical relevance to non-communicative diseases inflicting motor impairment.