2.6.1. Cross-Validation Scheme per Subject

We use a variation of grouped *k*-fold cross-validation per subject to evaluate the classifiers. We applied four-fold cross-validation by splitting the blocks of each subject into four continuous folds. Unlike regular cross-validation, we only used a single fold to train the classifiers while using the other three folds for validation. This scheme resulted in a training set of 9 blocks of 135 epochs each. We chose this approach since we are primarily interested in the performance of the classifiers in the case of low data availability. The classification task was to determine the cued target for each block. The fraction of correctly predicted cues provided the accuracy of a classifier. Data from all trials were used in the training stage, whereas classifier validation was performed multiple times per fold, each time using an increasing amount of trials (i.e., using the first trial, using the first two trials, etc., until all 15 trials have been used). For each of the 9 stimulated targets, the averages over the corresponding epochs across the utilized trials were used to predict the cued target in that block. The target with the maximum classifier score was then chosen as the predicted cued target. Before training the classifiers, a Z-score normalization transformation was developed on the training data to scale all EEG channels to unit variance. This transformation was then applied to the validation data.
