BCI based on Neural Correlates of Anticipatory Behavior during Driving

Anticipation of events such as, changes in traffic light signals and preparing to brake or accelerate are critical behaviors during driving. Smart vehicles, equipped with on-board Brain-Computer Interface (BCI), could decode the driver's intention to perform an action from his brain activity, thus enriching the interaction with its driver. To this end, the contribution of this thesis are three fold: (i) it presents 3 experiments to investigate anticipatory behavior from Electroencephalogram (EEG), while driving: count-down paradigm, traffic light changes in a virtual city and on a real road, (ii) it proposes methods for synchronous single-trial EEG classification of this behavior as well as asynchronous detection of movement intention, (iii) it explores new filtering techniques toward online application. In the first part, we present our first experiment, inspired from the classical Contingent Negative Variation (CNV) paradigm, where a count-down of numbers predicted the appearance of a cue that instructed to brake or accelerate accordingly. Through the EEG data (N=18), we show the presence of anticipatory potentials locked to the stimuli onset, which are similar to the well-known central negative Slow Cortical Potentials (SCPs). We further demonstrate the discrimination between cases requiring an action (brake/accelerate) upon an imperative subsequent stimulus (Start'/Stop' cues) vs. the events that do not require such action (count-down cues). We also show the possibility of detecting driver's movement intention through these potentials. In the second part, we extend the study to a more realistic scenarios using traffic lights through next two experiments. For the second experiment, we recorded 10 subjects over 3 days in a car simulator. During the second and third day, the subjects received online classification feedback together with reaction time after braking. Through this data, we confirm the presence of the anticipatory SCPs in response to the traffic lights as well as offline single trial performances with similar patterns to those of the first experiment. Interestingly, for the brake trials, we observed an improvement in the anticipatory behavior, which is likely due to the feedback provision. In the real car experiment on a closed road, we recorded EEG (N=8) over 2 days. Remarkably, we confirmed the existence of the anticipatory SCPs and demonstrated the possibility of detecting these potentials, despite large amounts of driving related visual distractions and movement artifacts. Thirdly, we report a post-hoc analysis on investigating the influence of filtering on the SCP detection performance. We present a new spectral filtering method, called Predictive Cascade Filter (PCF), which theoretically reduced the group delay associated with filtering of low frequency bands. The grand averages illustrate this reduction, whereas the classification performance did not improve by using the PCF filters. Indeed, the lowpass PCF as well as the usual lowpass filter appeared to be the best when applied causally (pertinent to online application), whereas the bandpass filter performed best, when applied non-causally (pertinent to offline analysis). We believe, the contributions presented in this thesis can impact the advancement of neuro-technology into smart vehicles as well as other applications such as neuro-rehabilitation.