Exploring the Impact of Data Continuity on EEG-Based BCI: A Comparative Study of AI Models

Abstract

This study conducted comparative experiments on continuous data for brain–computer interface (BCI) research using various models depending on the data composition. It aimed to compare the classification performance of artificial intelligence (AI) models on data that considered continuity, a characteristic of time series data, and data that did not. Specifically, this study elucidates data continuity as a characteristic of time-series data. Additionally, electroencephalogram (EEG) signals collected from five behavioral paradigms were categorized into six datasets based on continuity considerations. The experiment involved the training of AI models of similar complexity on the classified data to evaluate their classification performances. In this study, a classifier with interesting performance with respect to data continuity is found, and the results are discussed. In BCI research using EEG signals, action and thought classification achieved a high-performance score of 0.9071 in long short-term memory as data continuity decreased. Additionally, when continuity was not considered, deep neural network demonstrated maximum performance with a score of 0.9178. It was observed that data without consideration of continuity performed well in task classification. Therefore, BCI research focusing on behavior and accident classification based on EEG signals is expected to achieve excellent performance by showing various data characteristics through shuffling rather than considering data continuity. In conclusion, this study is expected to contribute to various research directions related to the continuity of time-series data in BCI research. In future research, the classification and case data will be diversified, and other models will be added to conduct performance comparison experiments using the data obtained based on feature extraction optimization.