Abstract
Multivariate classification analysis for non-invasively acquired neuroimaging data is a powerful tool in cognitive neuroscience research. However, an important constraint of such pattern classifiers is that they are restricted to predicting categorical variables (i.e. assigning trials to classes). Here, we present an alternative approach, Support Vector Regression (SVR), which uses single-trial neuroimaging (e.g., EEG or MEG) data to predict a continuous variable of interest such as response time, response force, or any kind of subjective rating (e.g., emotional state, confidence, etc.). We describe how SVR can be used, how it is implemented in the Decision Decoding Toolbox (DDTBOX), and how it has been used in previous research. We then report results from two simulation studies, designed to closely resemble real EEG data, in which we predicted a continuous variable of interest across a range of analysis parameters. In Simulation Study 1, we observed that SVR was effective for analysis windows ranging from 2 ms – 100 ms, and that it was relatively unaffected by temporal averaging. In Simulation Study 2, we showed that prediction was still successful when only a small number of channels encoded information about the output variable, and that it was robust to temporal jitter regarding when that information was present in the EEG. Finally, we reanalysed a previously published dataset of similar size and observed highly comparable results in real EEG data. We conclude that linear SVR is a powerful tool for the investigation of single-trial EEG data in relation to continuous and more nuanced variables, which are not well-captured using classification approaches requiring distinct classes.
Competing Interest Statement
The authors have declared no competing interest.
