My research focuses on the use of computational and machine learning techniques to study neural functions of the human brain, in healthy and pathological conditions. The main axes of my research lay in three areas:
In our day-to-day lives, we are constantly immersed in streams of sensory information. Often, stimuli from our environment, like sounds, do not occur in a random way, but follow statistical rules and repetitive patterns. Thanks to these rules we are able to learn regularities and form expectations about future events, before these occur. In my research, I am studying the neural and computational mechanisms that allow us to detect violations of anticipated events, in relation to our levels of consciousness, attention, or complexity of these events.
During my PhD I have been using machine learning techniques to study auditory processing in post-anoxic comatose patients. In particular, I have applied multivariate EEG decoding techniques to analyze EEG responses to auditory regularities at the single patient level. We were able to show that even patients who are deeply unconscious can detect violations of simple rules and complex sound sequences at a neural level. Importantly, the progression of auditory discrimination, quantified by the level of decoding performance over the two first days of coma, was predictive of the patients’ chances of awakening, allowing us to establish a clinical predictor of the patients’ outcome:
Tzovara et al., 2013, Brain, doi:10.1093/brain/aws264
Studying neural responses to environmental stimuli through electroencephalography (EEG), typically requires averaging hundrents or thousands of single-trial responses, and contrasting them at single electrode locations. During my PhD, I used multivariate techniques to model the distribution of single-trial EEG responses across the scalp, and extract topographic EEG responses in a data-driven way. These techniques can be used to model data at the single-patient level, to decode decisions from EEG responses, or to test the role of temporal intervals in processing environmental stimuli.
Tzovara et al., 2012, Develop. Neuropsychology, doi: 110.1080/87565641.2011.636851
Using environmental cues to predict threat is a crucial skill, encountered in many species, however a computational understanding of this process at a system’s level is lacking. We used modeling of psychophysiological responses (i.e. skin conductance and pupil size responses) to study which computational model best predicts activity of the human autonomic nervous system (ANS) during fear conditioning. We found that ANS activity is best explained by a probabilistic learning model that accounts for uncertainty in threat estimation. Furthermore, skin conductance and pupil responses map onto different learning quantities: pupil size reflects the estimated threat probability, but skin conductance reflects a combination of this and the uncertainty of the estimate.