Single Trial P300 Detection Using Dimensionality Reduction and Extreme Learning Machine

Brain-Computer Interfaces (BCI) are systems that function as a communication channel between people and external devices through brain information. P300 is an Event-Related Potential (ERP) widely used to decode an individual’s intent in BCI system applications, such as spellers in communication skill rehabilitation systems. However, the Signal-to-Noise Ratio (SNR), which quantifies a relation between the power of the transmitted signal and the power of the noise that corrupts it, for P300 is low so which makes it difficult to detect the using a single trial, which reduces the accuracy of the system. Additionally, in real-time BCI systems, it is necessary to maintain a good ratio of detection and execution time. In this work, three methods based on mean-based feature dimensionality reduction (PCA – Principal Component Analysis, FA – Factor Analysis, and MDS – Multi-Dimensional Scaling) as well as ELM – Extreme Learning Machine are presented, which allow efficient detection of a single trial P300 using 65.25% of the features normally used in the literature. PCA and MDS techniques reach an Area Under the Received Operation Curve (AUC) maximum of 0.90, and an average of 0.72, which are significant compared to the standard method based on the signal mean. Additionally, training times below 0.05 s are obtained, which are very important for real-time operation. The results allow us to conclude that the proposed methods are suitable for the detection of the single-trial P300, which can be used in BCI systems with real-time speller for rehabilitation engineering.