Echo State Networks Ensemble for SSVEP Dynamical Online Detection

Abstract

BackgroundRecent years have witnessed an increased interest in the use of steady state visual evoked potentials (SSVEPs) in brain computer interfaces (BCI), SSVEP is considered a stationary brain process that appears when gazing at a stimulation light source.New MethodsThe complex nature of brain processes advocates for non-linear EEG analysis techniques. In this work we explore the use of an Echo State Networks (ESN) based architecture for dynamical SSVEP detection.ResultsWhen simulating a 6-degrees of freedom BCI system, an information transfer rate of 49bits/min was achieved. Detection accuracy proved to be similar for observation windows ranging from 0.5 to 4 seconds.Comparison with existing methodsSSVEP detection performance has been compared to standard canonical correlation analysis (CCA). CCA achieved a maximum information transfer rate of 21 bits/minute. In this case detection accuracy increased along with the observation window lengthConclusionsAccording to here presented results ESN outperforms standard canonical correlation and has proved to require shorter observation time windows. However ESN and CCA approaches delivered diverse classification accuracies at subject level for various stimulation frequencies, proving to be complementary methods. A possible explanation of these results may be the occurrence of evoked responses of different nature, which are then detected by different approaches. While reservoir computing methods are able to detect complex dynamical patterns and/or complex synchronization among EEG channels, CCA exclusively captures stationary patterns. Therefore, the ESN-based approach may be used to extend the definition of steady-state response, considered so far a stationary process.HighlightsWe present a novel SSVEP dynamical detection approach based on ESN.This is the first time ESNs are applied to SSVEP based BCI systems.We provide experimental validation of proposed methodology.Experimental results indicate non-stationarity in SSVEP patterns.

Hypoarousal non-stationary ADHD biomarker based on echo-state networks

“…Reservoir-computing methodologies, and in particular ESNs, have proved to be an effective technique for EEG feature extraction [39]. Although ESNs have been applied to other EEG analysis scenarios including author's work in Parkinson's disease prognosis [37], this is the first time, to our best knowledge, that it has been used in the ADHD field.…”

“…Recordings are first filtered using a finite impulse response filter (FIR) at the following bands: theta1 (4-6 Hz), theta2 (6 -10 Hz), alpha1 (8 -11 Hz), alpha2 (10 -13 Hz), beta1 (13-20 Hz), beta2 (20-30 Hz), gamma1 (25-35 Hz) and gamma2 (35)(36)(37)(38)(39)(40)(41)(42)(43)(44)(45). The reference method has a substantial impact on potential measurements.…”

“…RC has been applied in the past to several EEG feature extraction and classification problems such as brain computer interfaces [34,35], epileptic seizure [36], prognosis in Parkinson's disease [37] or event detection [38]. In a previous work [39] we have demonstrated RC capacities to well characterize complex dynamics among EEG channels when a dominating frequency in the EEG spectrum is elicited. Here we explore EO-EC discrimination capabilities within EEG frequency bands, that as alpha may be characterized by a dominant oscillatory regime.…”