An Information Theoretic Approach for Dimensionality Reduction and Estimation of Network Modulations from High-Dimensional Electroencephalograms

• Activity Number: 673
• Type: Contributed
• Date/Time: Thursday, August 4, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Catherine Stamoulis*+ and Bernard S. Chang
• Companies: Harvard Medical School and Harvard Medical School
• Keywords: Information theory ; Dimension reduction ; Time series ; Electroencephalograms

Abstract

Electroencephalograms (EEG) measure aggregate neural activity and its dynamic modulations. In epilepsy, EEGs are routinely used to identify patient-specific seizure characteristics. For this purpose, patients are often monitored for several days, resulting in very large, high-dimensional and correlated data. Dimensionality reduction approaches that preserve the dynamic characteristics of information encoded in EEGs are desirable. Information theoretic measures are useful for reducing the high dimension of these signals and for quantifying neuronal network changes associated with impending seizures. In addition, these measures allow us to incorporate biologically-relevant dependencies, in terms of conditional probabilities, e.g., dependence of the correlation between EEGs on the history of their coordination, or the global connectivity sof the brain. We have developed sensitive and specific information theoretic parameters for quantifying neuronal modulations prior to seizure onset. We show that these parameters encode seizure-related changes in network coordination, strongly dependent on patient-specific baseline activity, and may ultimately be useful for seizure prediction.