Hypothesis-Driven Source Separation and Dimension Reduction of Neural Time Series Data

• Main Author: Lin, Hause Sihao
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01.01.2021
• Subjects: Engineering; Neurosciences; Psychology
• ISBN: 9798496548847

Neural dynamics spanning diverse spatial and temporal scales produce electrical fields that can be measured using non-invasive tools like electroencephalography (EEG), which records neural activity from multiple recording sites on the scalp of the head. However, most analytic approaches in EEG research do not fully exploit the rich, high-dimensional data to gain insights into how the brain processes information nor address the volume conduction problem. That is, EEG data contain a mixture of electric fields, produced by hundreds or thousands of neural sources, that propagate simultaneously to all recording sites on the scalp. Although this source mixing problem poses many analytic challenges, it also provides opportunities to develop better algorithms that can provide further theoretical and practical insights. Generalized eigendecomposition (GED) provides a multivariate framework for separating sources and reducing the dimensionality of EEG data while allowing for flexible hypothesis testing. I derived four analytic approaches from this framework and applied them to simulated and empirical data in four studies. Simulations in Study 1 showed that GED-based approaches were effective for separating sources, reducing data dimensionality, increasing signal-to-noise ratio, and testing hypotheses. Study 2 used GED to identify the determinants of cognitive control allocation and test the predictions of a theory of control allocation. Study 3 combined GED with time-frequency approaches to examine cognitive control processes during value-guided choice. Study 4 used GED to isolate sources whose activities reflected trial-varying decision attributes and showed that the timing of these activities was consistent with the predictions of decision models. In summary, my research provides evidence for the value of GED, a hypothesis-driven multivariate source separation and dimension reduction framework that can be used to improve our understanding of brain and cognitive function.