A General Aggregate Model for Improving Multi-class Brain-computer interface Systems' Performance

This paper proposes a general aggregate model for improving performance of multi-class Brain-Computer Interface (BCI) systems. In BCI systems, activation and delay are well known issues in conducting experiments. The delay of meaningful brain signal depends on subjects, tasks and experimental design. Therefore, within a trial it is not easy to identify where meaningful brain signal starts and ends. Most of current methods estimate the delay and extract a portion of meaningful brain signal in a trial and use this signal as a representative for the whole trial. Instead of doing so, our proposed aggregate model divides a trial into overlapping frames and treat them equally. These frames are classified and their results are then aggregated together to form classification result of the trial. From the general aggregate model, we derive two specific aggregate models using two stateof-the-art Common Spatial Patterns (CSP)-based methods for feature extraction. We performed experiments on Dataset 2a used in BCI Competition IV to evaluate the proposed models. This dataset was designed for motor imagery classification with 4 classes. Preliminary experimental results show that our proposed aggregate models are up to 8% better than the original CSP-based methods. Furthermore, we show that our aggregate model can be easily extended to online BCI systems