AbstractMind-wandering is a mental phenomenon where the internal thought process disengages from the external environment periodically. In the current study, we trained EEG classifiers using convolutional neural networks (CNN) to track mind-wandering. We transformed the input from raw EEG to band-frequency information (power), single-trial ERP patterns, and connectivity matrices between channels (based on inter-site phase clustering, ISPC). We used two independent EEG datasets for training and testing purposes (dataset A and B). In Experiment 1, we trained intra-subject models using the training dataset (dataset A) and compared the performance from three CNN architectures. We found that a simple CNN with two convolutional layers is sufficient. Accuracies varied across individual datasets and higher accuracies during validation were achieved with more unbalanced datasets, even though class size was balanced during training. In Experiment 2, we trained inter-subject models using dataset A and tested them on dataset B for an across-task prediction. Here, we used a stacking model to combine the outcomes of multiple input models. We achieved an accuracy of .68 during the across-task predictions, verifying the generalizability of our CNN. In both Experiment 1 and 2, the CNN performed best when trained on the raw EEG. Increased performance of the minority cases was found when training the models with a subset of more balanced datasets. Our study indicates the potential of training study-independent EEG classifiers with CNN using the raw EEG We also suggest that future EEG-machine learning studies should report the balancing status of the training dataset and the accuracies for each class.
Review for "Decoding kinetic features of hand motor preparation from single‐trial EEG using convolutional neural networks"