Convolutional Autoencoder-Based Dimensionality Reduction for EEG Microstate Analysis

Electroencephalography (EEG) microstate analysis is an established method for understanding brain dynamics by identifying quasi-stable states in neural activity. This study introduces a novel approach that integrates convolutional autoencoders (CAE) into the traditional microstate analysis pipeline, explicitly focusing on enhancing spatial pattern recognition. Our methodology employs a CAE architecture for dimensionality reduction of EEG topographic maps, followed by modified k-means clustering for microstate identification. The empirical evaluation reveals interesting dynamics in clustering performance, with the CAE approach showing improvements in certain metrics such as Silhouette Scores (0.3014 versus 0.2387) and Davies-Bouldin Index (1.2081 versus 1.4531), while also highlighting areas for future optimisation in cluster separation. This mixed performance provides valuable insights for the continued development of deep learning approaches in this domain. While maintaining essential neurophysiological features, our approach introduces robust deep learning capabilities as a foundation for future development of end-to-end deep clustering solutions in microstate analysis. This research contributes to the ongoing evolution of EEG microstate analysis by demonstrating how modern machine-learning techniques can be effectively integrated with established methods, while also identifying specific directions for further enhancement.