Performance of 1D-CNNs for EEG-Based Mental State Classification: Effects of Domain, Window Size and Electrode Montage

Deep learning paradigms have revolutionized the field of brain-computer interfacing and enabled the use of complex, nuanced methods for recognizing mental states. Prior work has demonstrated that these models can recognize the mental state in a variety of tasks, but few have specifically explored their performance concerning human factors that come into play with the use of brain-computer interfaces in day-to-day applications. For this research, we explored the use of 1D-convolutional neural networks to recognize two mental states–mental arithmetic and rest–from electroencephalograph signals. We focused our analysis on three parameters that affect the design and usability of brain-computer interfaces: input data representation (i.e. domain), window size (i.e. latency), and electrode montage (i.e. form-factor). In line with prior work, we found a clear bias in performance towards the frequency domain representation. We also found that training our model with short windows of time (i.e. 0.25s) provided close to peak accuracy. Furthermore, high accuracy was maintained with sparse electrode subsets of the full 10-20 system. We discuss these findings and how they can contribute to ongoing work to bring deep learning enabled brain-computer interfaces into day-to-day applications.