Effects of stimulus modality and response type on oddball stimulus discrimination using polarity-considered EEG microstate labeling

ObjectiveBrain-computer interfaces (BCIs) require effective feature extraction and dimensionality reduction from multidimensional brain signals. Electroencephalogram (EEG) microstate analysis offers a fast and noise-resistant approach by classifying the states of brain signals into spatial distribution patterns (templates). Each EEG segment was assigned the template with the highest spatial correlation, reducing the information to a one-dimensional representation. However, prior BCI studies have often ignored the polarity of spatial distributions in these templates. Incorporating polarity during labeling may enhance classification performance. This study investigated the effectiveness of polarity-considered microstate labeling for classifying infrequent stimuli in an auditory-visual oddball task with implications for BCI applications. MethodEEG recordings were analyzed using polarity-considered microstate labeling to classify infrequent stimuli. This study examined the effects of stimulus modality (auditory or visual), modality conditions (unimodal: stimulus and response in the same modality; cross-modal: stimulus and response in different modalities), and response type (key-press task vs. mental counting task) on classification accuracy. Machine learning models were used for classification, including support vector machine, random forest, logistic regression, XGBoost, CatBoost and K-means methods. ResultsPolarity-considered labeling outperformed the non-polarity approach, especially in decision-tree-based models (20.1% improvement in the key-press task and 22.2% improvement in the mental counting task). A significant interaction was observed between stimulus modality and response type, with the highest accuracy achieved when the infrequent stimuli in the key-press task involved cross-modal visual information. ConclusionThe findings suggest that polarity-considered microstate labeling enhances EEG-based classification. This approach has potential applications in BCI, such as in P300 spellers using cross-modal auditory-visual stimuli.