Multivariate Analysis Methods in Cognitive Neuroimaging: Applications in Basic and Clinical Research

Multivariate decoding methods have revolutionized cognitive neuroimaging in recent years by enabling the extraction of spatially distributed neuronal responses not accessible to traditionally used univariate methods. Importantly, in addition to providing increased analytical sensitivity, these new approaches have also led to fundamental neuroscientific paradigm shifts. The possibility to extract spatially distributed neuronal information for instance has fortified the proposition that cortical function is not organized in an exclusively modular fashion. The options to analyze individual subject data and quantify neuronal information content based on pattern dissimilarities have further paved the way for new neuroscientific questions. We make use of these new possibilities and combine multivariate methods, such as classifiers from Machine Learning (ML) or Representational Similarity Analysis (RSA), with traditional univariate approaches to address open questions within three research areas in cognitive neuroimaging. In the first research topic, we assess the potential of classification methods from ML to advance clinical diagnosis of psychiatric disorders based on multi-site resting-state functional magnetic resonance imaging (RS-fMRI) data. Conventional diagnostic approaches face a variety of challenges as they commonly require a team of specialists and a battery of behavioral tests. This approach is often costly and time consuming, with the potential to deliver ambiguous results. The ability of ML classifiers to extract distributed neuronal information from RS-fMRI data promises to advance the establishment of biomarkers underlying psychiatric disorders, since such disorders commonly affect cortical structure and function in a global manner. With Autism Spectrum Disorders as an example, we ask how accurately ML classifiers can make diagnostic predictions based on RS-fMRI data, assess a variety of auxiliary methods from ML which can increase diagnostic sensitivity, and investigate how remaining caveats can be addressed. In the second research topic, we assess the effect of expectation on somatosensory processing based on behavioral and fMRI data. It has been previously shown that expectation dampens the mean BOLD signal while concurrently improving behavioral dissociation performance. The underlying neuronal mechanisms leading to this seemingly contradictory effect are, however, still under debate. In accordance with previous studies, we ask if expectation recruits fewer neuronal units, that specialize in encoding the expected stimuli with high selectivity. This would imply increased signal-to-noise ratio in favor of the expected stimuli, and could explain the decrease in mean BOLD signal associated with behavioral improvements. Multivariate decoding methods allow us to empirically test this hypothesis, as an increase in signal-to-noise ratio should improve the decoding of neuronal activity patterns associated with expected stimuli. The decision outcome of classifiers can be matched to the behavioral decisions of individual subjects. This allows us to connect the two experimental levels for a systematic explanation of the observed effects induced by expectation. In a third research topic, we turn our attention to Brain-Computer Interface (BCI) technology, which promises to pave the way for assistive devices or for devices employed in therapeutic settings. In order to optimise future BCI technology, more reliable read-outs of neuronal states are necessary to provide a basis for the control of external devices. We ask if ML classifiers can reliably predict the up- and down-regulation of motor excitability based on individual subject electroencephalography (EEG) measurements. In a rst step, subjects learn within a neurofeedback paradigm to attain the desired motor states. This is accomplished with a new protocol, where subjects are trained to modulate their motor states through real-time visual feedback of motor evoked potential (MEP) amplitude in response to transcranial magnetic stimulation (TMS). Once subjects learn a robust regulation of their motor excitability, EEG data associated with both states is acquired. Based on these EEG measurements, we assess the capability of Machine Learning classifiers to predict the motor states. In addition, we employ feature selection techniques to assess which components of the EEG signal are most relevant for the prediction. Feature selection can aid robust prediction, and promises to elucidate underlying neuronal mechanisms associated with the up- and down-regulation of the motor system. The capability of Machine Learning methods to draw inference from single subject data and based on only a few trials promises to enable real-time BCIs tailored to the needs of individual subjects. In summary, the three research projects presented in this thesis show how the application of multivariate decoding analysis can lead to fundamental neuroscientific insights which provide a basis for practical applications and a better understanding of the mechanisms underlying sensory processing.