Investigating Various Approaches in Classification of EEG Signals Representing Distinct Cognitive States to Reach an Optimal Solution
MetadataShow full item record
There are various cases in which cognitive neuroscientists might be interested in exploring the neural differences associated with distinct cognitive states such as whether an individual has remembered some information or not. While it is common to use event-related potentials (ERPs) to distinguish neural activities representing different cognitive states, it does not allow us to explore single events because of its averaging nature. Classification of brain states associated with single events using real-time signals holds great potential for real-world applications such as brain-computer intervention systems that could support everyday learning. However, the progress in reaching high classification accuracy is still in early stages and thus, moving to the next step and creating such interventions is not possible yet. Moreover, previous studies applying classification methods to decode cognitive states have not typically compared different methods or explained the reasons for their choices. As a result, in this study, I systematically compared different methods of feature extraction, feature selection, and choice of classifier in the same study to investigate which methods work the best for decoding different episodic memory and perceptual “brain states.” Using an adult lifespan sample EEG dataset collected during encoding and retrieval of objects paired with color and scene contexts, I found that the Common Spatial Pattern (CSP)-based features could distinguish the trials of different memory classes (i.e. item remembered vs. forgotten; context correct vs. incorrect; red vs. green vs. brown context perception) better than other types of features (i.e., mean, variance, correlation, features based on AR model, and entropy), and the combination of filtering and sequential forward selection was the optimal method to select the effective features. Moreover, Bayesian classification performed better than other commonly used options (i.e., logistic regression, SVM, and LASSO). These methods were shown to outperform alternative approaches for an orthogonal dataset, supporting their generalizability. My systematic comparative analyses allow me to offer some recommendations for cognitive researchers to consider when applying machine learning based classification to their datasets.