Idle State Detection with an Autoregressive Multiple Model Probabilistic Framework in SSVEP-Based Brain-Computer Interfaces

Abstract

The detection of the idle state is a key feature in developing asynchronous steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs). Despite the large number of successful SSVEP detection methods, only a few studies have explicitly included the detection of the idle state. This work demonstrates the feasibility of a novel autoregressive multiple model (AR-MM) probabilistic framework for the detection of SSVEPs and the idle state. In a MM framework an SSVEP is identified by selecting one of the candidate models, each representing a particular SSVEP class, that best represents the dynamics of the data. An average classification accuracy of 78.94 ± 10.28% and an information transfer rate (ITR) of 28.85 ± 9.39 bpm are obtained for the 6-class SSVEP dataset in a longitudinal study. Furthermore, this work quantifies the performance of the AR-MM framework, that provides a measure of probability, for idle state detection. An average area under curve (AUC) of 0.83 is achieved with different threshold settings for idle state detection. The idle state could be detected with 81% average accuracy when considering maximum idle state and non-idle state detection rates. The results, obtained from a single-channel analysis, validate that the AR-MM framework is a good candidate for SSVEP detection and also for the idle state detection when compared with two multivariate methods, the canonical correlation analysis (CCA) and its extension, the filter bank canonical correlation analysis (FBCCA). With only a pair of electrodes required for the AR-MM approach, this is more practical for daily use of BCI applications, where a minimal amount of channels are desirable.