Optimizing parameters and algorithms of multivariate pattern classification for hypothesis testing in high-density EEG

Abstract

Multivariate pattern analysis (MVPA) has come into widespread use for analysis of neuroimaging data in recent years and is gaining further momentum. Given the task of detecting a generalizable pattern in neural activity, MVPA allows to detect fine multidimensional spatiotemporal contrasts between two or more conditions and is thus able to take the full advantage of multivariate information encoded in the data. In particular, MVPA based approaches lend themselves very well to the analysis of electroencephalogram (EEG) data because, unlike the widely-used averaging methods, they consider the signal in its entirety and are thus less susceptible to the confounding effects of single points with abnormal amplitudes. However, using MVPA for hypothesis testing purposes in high-density EEG data has remained a challenging issue. Although MVPA is getting more and more mainstream to detect information in neural activity, its behavior is not well understood, yet. EEG data are high dimensional, yet sample size is usually low in comparison. Moreover, due to the low signal-to-noise ratio, the effect size is small and differences between classes are hard to detect. In such cases, MVPA behaves unexpectedly which makes the overall accuracy of the classifier difficult to interpret. In addition, because MVPA is sensitive to any kind of structure in the data, confounding factors or additional variance within data can bias accuracy. Such complexities warrant extra caution when interpreting classification results, thereby requiring further investigation and guidelines. On the other hand, MVPA literature is mainly dominated by methods suited for fMRI data and most of the dedicated EEG methodology is developed for brain computer interfaces (BCI) or single trial analysis of event-related potentials. Specifically, decoding continuous EEG increasingly suffers from the curse of dimensionality because of the lack of clear prior knowledge on which frequency bands and time points carry relevant information, or an onset where the effect of stimulation can be expected. In this thesis, we addressed the aforementioned challenges involved in using MVPA for decoding EEG data. Chapter 2 describes the statistical properties of MVPA in realistic neuroimaging data and provides important guidelines to interpret classification results. We show that the probability distribution of classification accuracies does not follow any known parametric distribution and can be strongly biased and skewed. We describe unexpected properties of the distribution of classification rates which forbid their use as estimates of the size of experimental effects. Importantly, we scrutinize the finding of below chance level classification rates, which often occur in low sample size, low effect size data and their implications on the shape of classification rates distribution. Next, in chapter 3, we investigate neuroimaging data that, next to a main effect of class, additionally contains a nested subclass structure. We show that in these data sets, correct classification ratios are systematically biased from chance even in absence of class effect. We propose a nonparametric permutation algorithm which can detect the subclass bias and account for its effect by adjusting permutation tests to consider the subclass structure of the data, using subclass-level randomization. Finally, in chapter 4, we used MVPA to decode continuous high-density EEG across subjects. We developed a classification framework along with a specific preprocessing procedure that is optimized for three purposes: 1) to increase signal-to-noise ratio, 2) to reduce the dimensionality of the data, and 3) to adapt the signal better to between-subject classification. Our algorithm uses a two-step classification procedure based on ensemble of linear support vector machines (SVM) which learns the spatial and temporal components of neural activity separately and then aggregates the two components of information to build a classification hyperplane using another linear SVM. We then use this method to see whether human sleep EEG contains any information about what has been learned before sleep.