The correlation-based model: An alternative system for analyzing ERP data in cognitive research

Abstract

This study is focused on the development of a model for analyzing electrophysiological data (EEG) utilizing the evoked potentials (ERP) method used in cognitive research. The aim of the model is to overcome several limitations arising from traditional methods of ERP analysis. The model was tested for its ability to distinguish between dyslexic and regular readers. ERP data collected during a typical experiment contain a large amount of information that is not utilized during data analysis. For instance, it is acceptable to define a component such as the P300 according to the peak of a wave based on a few electrodes. Furthermore, this is often accomplished based on the researcher's subjective impression. Information such as the pattern of the wave, its width, rate of ascent, rate of descent, and the impact of the stimulus that evoked it over the entire scalp, is left out. In contrast, the modeling method proposed here considers all available information, utilizing a precise algorithm that allows fully automated analysis. The method produces a typical profile for a given type of subject and can determine for each new subject his/her similarity to that type. The basic idea behind the model is that two subjects exposed to the same task would share certain similarities in their electrophysiological data. However, the latency of this shared similarity might differ from person to person. It is thus possible to take the data from one subject and systematically seek the point of maximal similarity across all electrodes. Consequently, the data matrix of subject A in the area of P300 (e.g., all information on all electrodes between 250 and 350 ms) will be correlated with the data of subject B of the same width, i.e., 100 ms running from 200 ms throughout 400 ms. The correlation between the two matrices is expected to rise as the area of similarity gets closer, and descend once past that point. The size of the maximal correlation indicates the similarity between the two subjects, and its location in time indicates differential latencies. The averaging of data across subjects should be carried out at the points of maximal similarity. In order to demonstrate the power of the method, this study constructed two separate models for a dyslexic and regular reader, respectively. The models were based on ERP data recorded during a lexical decision task. The similarity of each subject to both models was calculated. This method correctly classified 68% of subjects. In addition, the model was able to detect sub-groups within each diagnostic category with distinct behavioral patterns.