A comparison of temporal windowing schemes for single-trial ERP detection

Single trial ERP detection is critical for stimulus-synchronous brain computer interfaces. This paper presents a comparison of three different algorithmic schemes for single-trial ERP detection: SVM (baseline), hierarchical SVM-(naive) Bayes, selected temporal windows-based SVM-(naive) Bayes. An ERP-based image search system, including experimental setup, data collection, pre-processing, and three ERP detection schemes is described and utilized as the framework for comparison. We apply three schemes on EEG data from four subjects acquired on four days (eight sessions) each. Results indicate that a properly trained SVM operating on data from the post-stimulus [0,500]ms interval and SVMs trained on 50 ms nonoverlapping windows spanning the poststimulus [200,450]ms interval (where P300 is expected) whose binary decisions are fused via the naive-Bayes approach perform similarly in terms of area under the ROC curve measure, while the latter fusion approach applied to all ten nonoverlapping windows spanning the [0,500]ms poststimulus interval is inferior. The poststimulus time limit of 500 ms is imposed on all data that goes into the ERP detector because in our experimental setup the subjects are asked to press a button when they recognize the event of interest, which creates motor responses in the brain typically in the [600,800]ms interval and around.