Enhancing cross-subject transfer performance for SSVEP identification using small data-based transferability evaluation

Abstract

Cross-subject steady-state visual evoked potential-based brain-computer interfaces (SSVEP-BCIs) have attracted increasing attention in recent years due to their potential to reduce calibration efforts and improve usability. This study proposed an efficient cross-subject transfer method for SSVEP-BCI, which significantly reduced the calibration effort and enhanced system performance. By employing a minimal dataset obtained from the target subject, specifically, a single trial for each frequency, we calculated the similarity between the SSVEPs of the target subject and the source subjects. Based on the similarities, a transferability evaluation model was constructed, which was then used to assign weights to the source subjects. The features calculated from the source subjects were then weighted and summed to obtain the frequency recognition results. The proposed cross-subject transfer method was evaluated in two pipelines based on the ensemble task-related component analysis (eTRCA) and task-discriminant component analysis (TDCA), respectively. The proposed method in eTRCA-pipeline initially achieving accuracies of 96.98 ± 3.66 % and 88.97 ± 14.66 % in our dataset and the Benchmark dataset, respectively, at a data length of 1 s. After fusing the features of eTRCA with those of the extended canonical correlation analysis (eCCA), the method exhibited enhanced performance, elevating accuracies to 98.23 ± 2.68 % and 92.83 ± 10.75 %. Similarly, in TDCA-pipeline, accuracies initially reached 97.47 ± 3.88 % and 93.53 ± 10.22 % in our dataset and the Benchmark dataset, respectively, and after fusing TDCA features with eCCA features, they remarkably increased to 98.13 ± 2.89 % and 95.00 ± 8.43 %, respectively. This study offers a promising solution for the fast and effective calibration of SSVEP-BCI, paving the way for more practical and accessible BCI systems.