Methods for Multivariate Time-Series Classification on Brain Data : Aggregation, Stratification and Neural Network Models

Abstract

This thesis is about the analysis of two data sets consisting of human brain data measured by electroencephalography (EEG). One data set contains data from craving smokers, who have not smoked for several hours, non-craving smokers, who had a smoke shortly before the measurement and non-smokers. These classes are to be distinguished with the help of neural networks. The second data set contains noisy EEG signals with different kinds of noise from and clean signal that are to be distinguished. In order to analyze them, I adapt a network structure, that was originally developed for neural networks for object recognition in images. I modify, the so called residual blocks in order to use them on EEG time series. One difficulty of EEG data is their property of being individual-specific. This can sometimes even be helpful to get improved predictions: if the classes of the data change so infrequently that it can be assumed that several parts (so called snippets) of a longer signal belong to the same class, then this information can be used to make predictions of several snippets and aggregate them to create a classification of the longer original signal.<br />I investigate a total of 15 research questions, regarding the context in neuroscience, adaptations and improvements of neural network models, and the optimal choice of aggregation functions.