Consensus clustering for group-level analysis of event-related potential data

Understanding human brain activity through spatiotemporal electroencephalogram (EEG) analysis has gained prominence, with cluster analysis emerging as a valuable tool. While traditional event-related potential (ERP) analysis techniques for identifying interesting ERPs involve subjective time window selection, conventional cluster analysis focusing on spatial dynamics amplifies the risk of component identification errors when data is imperfect. Consequently, they do not offer a unified, appropriate time window determination approach for testing experimental hypotheses. This thesis introduces a series of consensus clustering-based approaches for examining brain responses in spatiotemporal ERP/EEG data. Specifically, the first study proposed a data-driven approach for determining the optimal number of clusters by evaluating the inner similarity of the estimated time window. A consensus clustering method from diverse clustering methods was also designed, including an M-N plot method for configuration. The second study proposed a multi-set consensus clustering approach across individual subjects to determine an appropriate (i.e., precise and stable) time window of ERP of interest. The time window determination method we developed examined two criteria for selecting a representative cluster map: inner similarity and hypothetical temporal coverage. The third study presented a multi-set consensus clustering approach for clustering analysis of single-trial EEG epochs that aimed to identify individual subjects’ evoked responses (ERP components). This study also introduced a standardized approach for evaluating scores from signal processing methods. Lastly, the fourth study introduced an ensemble deep clustering pipeline for reliably determining the time window when data quality is imperfect, revealing the adeptness of deep neural networks in feature extraction and time window determination. In conclusion, this thesis offers a promising computational framework for ERP identification in group-level analysis. The aforementioned studies enhance our understanding of human brain function, have broad implications for computational neuroscience, and suggest adaptable solutions for future neuroimaging investigations. Keywords: Electroencephalography (EEG), Event-related potentials (ERPs), ensemble learning, consensus clustering, time window, cognitive process, deep clustering, cluster aggregation.