Ensemble strategies exploration for the calibration data optimized spatial filters based SSVEP recognition algorithms

Abstract

Steady-state visual evoked potential (SSVEP) based brain-computer interfaces (BCIs) exhibit high information transfer rate (ITR) characteristics. However, the electroencephalogram (EEG) data encoding SSVEP patterns has a low signal-to-noise ratio, leading to limitations in applications. To enhance the application of SSVEP-BCIs, various SSVEP recognition algorithms have been developed by utilizing prior knowledge based on the calibration data optimized spatial filters (CDOSF). Yet, the process of combining spatial filters from multiple stimuli remains unexplored, resulting in performance bottlenecks for SSVEP recognition and application scenarios. In this study, we provide an overview of four popular CDOSF-based SSVEP recognition algorithms and introduce three ensemble strategies to extend these recognition algorithms. For comparison purposes, we selected three benchmark SSVEP datasets and conducted comparative experiments on the three ensemble strategies across the four recognition algorithms. The experimental results demonstrate that the ensemble strategies significantly enhance SSVEP recognition accuracy and ITR. Ablation studies also confirm that ensemble strategies possess characteristics of requiring less calibration data, fewer EEG channels, and insensitivity to parameters. The ensemble strategies extended CDOSF-based recognition algorithms offer a novel choice for constructing SSVEP-BCIs.