A novel hybrid method based on task-related component and canonical correlation analyses (H-TRCCA) for enhancing SSVEP recognition

TL;DR

This paper introduces a new method for improving brain-computer interfaces by combining two techniques to better recognize brain signals with minimal training.

Contribution

A novel hybrid method (H-TRCCA) is proposed, combining task-related component and canonical correlation analyses for SSVEP recognition with limited calibration data.

Findings

H-TRCCA achieved 91.44% accuracy with only two training trials per frequency in Dataset I.

The method outperformed existing techniques using fewer training trials while maintaining high information transfer rates.

It showed robust performance with maximum average information transfer rates of 188.36 bits/min and 139.96 bits/min for two datasets.

Abstract

Brain-computer interfaces (BCIs) based on steady-state visual evoked potentials (SSVEP) rely on the brain’s response to visual stimuli. However, accurately recognizing target frequencies using training-based methods remains challenging due to the time-consuming calibration sessions required by subject-specific training methods.

To address this limitation, this study proposes a novel hybrid method called Hybrid task-related component and canonical correlation analysis (H-TRCCA). In the training phase, four spatial filters are derived using canonical correlation analysis (CCA) to maximize the correlation between the training data and reference signals. Additionally, a spatial filter is also computed using task-related component analysis (TRCA). In the test phase, correlation coefficients obtained from the CCA method are clustered using the k-means++ clustering algorithm. The cluster with the highest average correlation identifies the candidate stimuli. Finally, for each candidate, the correlation values are summed and combined with the TRCA-based correlation coefficients.

The H-TRCCA algorithm was validated using two publicly available benchmark datasets. Experimental results using only two training trials per frequency with 1s data length showed that H-TRCCA achieved average accuracies of 91.44% for Dataset I and 80.46% for Dataset II. Additionally, it achieved maximum average information transfer rates of 188.36 bits/min and 139.96 bits/min for Dataset I and II, respectively.

Remarkably H-TRCCA achieves comparable performance to other methods that require five trials, utilizing only two or three training trials. The proposed H-TRCCA method outperforms state-of-the-art techniques, showing superior performance and robustness with limited calibration data.

Figures