Abstract:
All of the experimental and theoretical work presented in this dissertation has been inspired by the general idea of applying event-related brain potential (ERP) measurement and assessment for practical purposes: cognitive diagnostics and Brain-Computer Interfaces (BCI) for paralyzed people.
In Chapter 1, two new ERP paradigms are introduced, which were developed for the diagnostics of a particular cognitive function, the recognition of affective prosody. The affective state of a speaker can be identified from the prosody of her/his speech. Voice quality is the most important prosodic cue for emotion recognition from short verbal utterances and nonverbal exclamations, the latter conveying pure emotion, void of all semantic meaning. Two context violation ERP paradigms, passive oddball and priming, were adopted for the study the ERPs reflecting this recognition process. A new negative wave, the N300, was found in the ERPs to contextually incongruous exclamations. This component was interpreted as analogical to the well known N400 response to semantically inappropriate words. The N300 appears to be a real-time psychophysiological measure of spontaneous emotion recognition from vocal cues, which could prove a useful tool for the examination of affective-prosody comprehension for the purposes of both fundamental psychological research and applied clinical diagnostics.
Chapters 2 and 3 address the important issue of ERP component detection and quantification mostly from the perspective of individual data assessment, which is crucial for any reliable ERP diagnostics. The classical measures, area and peak, used in Chapter 1 are the most popular estimators of ERP components. Although they are in standard use in ERP research, they are very rough estimators, which heavily rely on visual inspection of the waveforms, and are thus prone to experimenter bias. Chapter 2 shows how the Continuous Wavelet Transform (CWT) can be used in ERP data analyses. A novel assessment method, the total-average-CWT, is introduced and demonstrated on the ERP data acquired in the emotional prosody experiments presented in Chapter 1. This method does not rely on visual inspection of the waveforms but allows for more automated detection of peaks. At the same time it provides more precise estimation of ERP components and yields larger statistical difference effects than classical methods do.
The usage of the ERP technique in clinical applications for diagnostic purposes, requires special methods of EEG data assessment, based on single trial analysis. The total-average-CWT method introduced in Chapter 2 is an example for one such method. In Chapter 3, a new, largely improved version of the method, the t-CWT, is introduces. The t-CWT is based on the CWT and Student's t-statistics. The method was systematically tested in two prototypical ERP paradigms, oddball and sematnic priming, which belong to the basic tools of ERP-based cognitive diagnostics. The method was compared to other assessment procedures based on Principal Component Analysis (PCA) and the Discrete Wavelet Transform (DWT). Similarity to clinical settings was achieved by the individual assessment of each participants ERP data. Both whole waveforms and single ERP components were assessed by multivariate procedures including PCA data set reduction, Hotelling's test and a randomization test. The assessment of the whole ERP waveforms is particularly relevant to the paradigms of the context violation class introduced in Chapter 1. The results demonstrated the superiority of the t-CWT to the other assessment methods.
In the study presented in Chapter 4, the detection and quantification method introduced in Chapter 3, the t-CWT, was applied in the classification of single ERP trials for the purposes of BCI. In this application, the t-CWT is used as a general feature extraction method, which provides the optimal variables describing the pattern that best discriminates between the ERPs
reflecting different cognitive processes. The method has been validated in the International BCI Competition 2003, where it was a winner (provided best classification) on two ERP data sets acquired in two different BCI paradigms, P300 speller and Slow Cortical Potential (SCP) feedback. In the P300 speller paradigm the method provided results, which were as good as those obtained from simple and redundant features with a very powerful classifier based on machine learning, the Support Vector Machines (SVM). The t-CWT method has the advantage that it is very simple, intuitively plausible, readily visualizable and the extracted features have clear interpretation as ERP components.
To summarize, two major results are presented in this dissertation: first, the newly found ERP component, the N300 to inconsistent affective prosody expressed in emotional exclamations, and second, the newly developed t-CWT feature extraction method for fully automated detection and quantification of ERP components.