Accurate estimation of information transfer rate based on symbol occurrence probability in brain-computer interfaces

Abstract

ObjectiveThe common criteria for evaluating the performance of the brain–computer interface (BCI) are classification accuracy and information transfer rate (ITR). Due to the fact that the BCI system has direct interaction with patients, the accuracy of estimated ITR is very influential. Although the most popular method for ITR estimation is the Wolpaw’s definition, the estimated ITR by this definition is often inaccurate in online applications. One of the existing limitations of it is that all symbols supposed to have the same occurrence probabilities, but symbols do not share the same probability in most real-world applications. MethodsIn this paper, a comprehensive ITR formula is proposed based on symbol probabilities, using the general concept of mutual information. ResultsThe Wolpaw’s definition leads to a strong ITR over-estimation compared to considering the symbol real occurrence probabilities. This estimation error increases with increasing classification accuracy and the number of symbols. For shorter required time to select one symbol, the ITR estimation error is also greater. ConclusionThe proposed Method estimates the ITR more accurately in online applications. SignificanceThe presented formulas provide simplified ITR definition based on symbol probabilities corresponding to a variety of BCI hierarchical structures.