Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Modeling single-trial event-related potentials using hidden multivariate patterns
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Gabriel Weindel1, Jelmer Borst2, Audrey Bürki1; 1Universite de Lausanne, Switzerland, 2University of Groningen, The Netherlands
In reaction-time-based experimental tasks, a repeated sequence of neural events typically occurs between the onset of a stimulus and the production of a response. To characterize these events, electroencephalographic recordings are often analyzed using event-related potentials (ERPs). ERPs are computed by averaging many trials in the hope that task-related activity is time-locked to the stimulus (or the response), while EEG task-unrelated activity is canceled out in the average. A critical assumption of this type of analysis is that the time variation across trials of these task-related neural events is relatively constant. However, reaction time modeling and single-cell measurements have shown that these trial-by-trial time variations, even for low-level visual components, can be much higher than desired for ERP analyses. In the best case, this results in too smooth ERP estimates, in the worst case to missing neural events entirely. To address this time jitter problem in ERP research, we present the hidden multivariate pattern method, which estimates single-trial ERPs by taking into account information present in all trials. The hidden multivariate pattern method (HMP) relies on the assumption of a time-resolved sequence of neural events leading to a response by modeling patterns that recur in the stimulus-response interval across trials. The method assumes that single-trial ERPs originate from a pattern (e.g., a half-sine) with a multivariate representation across all scalp electrodes. These patterns are expected to capture task-relevant cognitive processes that are sequential due to the nature of the task (e.g., stimulus encoding, decision-making, and response execution), but whose timing varies across trials. The sequential constraint allows us to regularize the parameter space that the algorithm must explore to infer the timing of these events in a single trial. The method is more than just an ERP time-correction algorithm, as it is designed to infer the number of events contained in a task or condition and to provide an estimation of the time location of each component at the single-trial level. We will demonstrate how estimates from this method inform us about the cognitive processes involved in various tasks. First, using a decision-making dataset, we show how the method parses the EEG into sequences of visual encoding, attention orientation, decision-making, and motor execution events. This decomposition reveals that different periods in the reaction time, captured by the time interval between different events, are sensitive to different aspects of the stimuli and the task instructions. It also shows that we can apply a cognitive model of the task to the intervals between events using the single-trial estimates. Then, in a cognitive control dataset, we show that, in addition to decision-making events, we can identify conflict-specific events whose timing aligns with the theoretical expectations of the temporal course of conflict between task-relevant and task-irrelevant stimulus features. Finally, using public datasets, we show that language ERPs such as the N400 and the P600 can be described at the single-trial level using HMP and that their timing provides insight into the time course of lexical processing.
Topic Areas: Methods, Control, Selection, and Executive Processes