Computational methods for EEG recordings during periodic sensory stimulation

Understanding how the human brain processes sensory stimuli remains challenging. For this purpose, brain responses elicited by brief stimuli and recorded with scalp electroencephalography (EEG) have been widely considered. Besides, much less is known about the dynamics of EEG responses and subsequent perception evoked by longer-lasting, tonic stimulation. Whereas phasic responses can be extracted from background noise thanks to averages across trials, it is far more difficult to highlight tonic responses without any temporal cues along long time intervals. In this context, this thesis considers periodic sensory stimulation to better link brain responses to input tonic stimuli. This offers several advantages, as a periodic stimulus elicits periodic activity in neuronal populations, referred to as a steady-state response (SSR). The known periodicity conveniently enables highlighting the dynamics of the elicited responses. In this dissertation, this approach is designed and implemented to study thermosensation, a vital sense to enable proper thermoregulation, to identify textures and to assess threats arising from noxious thermal stimuli. The EEG responses and perception elicited by warm and cool stimuli whose intensities are slowly and periodically modulated are recorded and analyzed. The baseband EEG responses as well as ongoing oscillations in physiological frequency bands are comprehensively studied. Despite the advantages of periodic stimulation, their benefits remain conditioned to a sufficient signal-to-noise ratio (SNR) of the analyzed responses. Unfortunately, EEG recordings typically exhibit a very low SNR, harming their direct exploitation in numerous settings. For this reason, this thesis proposes and compares methods to extract periodic components from multidimensional recordings. These approaches consist in (1) spatial filters maximizing the periodicity of the extracted components, (2) spatial filters optimizing the specificity of such periodic components to classes of multidimensional signals and (3) tensor factorization schemes to highlight the most periodic components through unsupervised algorithms. Simulations illustrate how the SNR of SSRs can be significantly enhanced to facilitate the use of periodic stimulation in practical contexts.