High-performance Deep Neural Network Pretrained with Contrastive Learning for Asynchronous High-frequency c-VEP Detection

In order to reduce the visual fatigue during use, a high-frequency discrete-interval binary sequence (DIBS) was proposed for an asynchronous 4-target code-modulated visual evoked potential (c-VEP) brain-computer interface system. However, with traditional spatial filter-based decoding methods, some of the subjects have difficulties activating the high-frequency system from the idle states, which indicates that the system's effectiveness declined because of the user-specificity. A deep neural network was therefore built, consisting of two-way LSTM networks, for enhancing the decoding performances in the high-frequency stimulus cases. The architecture includes a pretraining phase with contrastive learning, a training phase and a fine-tuning phase. In the pseudo-online experiments, compared to the results of the usual filter-bank task-related component analysis (FB-TRCA) method, the invalid trial percentage was reduced from 22% to zero, the average reaction time (RT) from 2.75 s to 1.39 s, and the average false positive rate (FPR) from 8.33 × 10 ^–2 min ^–1 to 1.39 × 10 ^–2 min ^–1 by using the designed architecture. In this study, the detection of high-frequency c-VEP has been greatly improved, and the response signals were proved to contain useful information. All of the subjects were able to activate the system, which also verified the feasibility of high-frequency stimuli in other related experiments in the domain.