Representative-Based Cold Start for Adaptive SSVEP-BCI.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Human-Computer Interaction

Background:

• Steady-state visual evoked brain-computer interfaces (SSVEP-BCI) offer a promising avenue for assistive technology.
• However, extensive calibration requirements often impede user experience and widespread adoption.
• Existing methods struggle with model generalizability across different users, devices, and sessions.

Purpose of the Study:

• To develop a cross-dataset adaptive model for SSVEP-BCI that minimizes calibration effort.
• To improve model generalizability and maintain high prediction performance without user-specific training.
• To enable a more plug-and-play and high-performance BCI experience.

Main Methods:

• A user-independent (UI) representative model is selected from a data pool for new users.
• Online adaptation and transfer learning techniques are applied using limited user-dependent (UD) data.
• The method was validated through offline (N=55) and online (N=12) experiments.

Main Results:

• The representative model approach reduced calibration by approximately 160 trials compared to UD adaptation.
• Online experiments showed a reduced time window (0.56±0.2s) with high prediction accuracy (0.89-0.96).
• The highest reported information transfer rate (ITR) of 243.49 bits/min was achieved in a calibration-free setting.

Conclusions:

• Representative UI models can be effectively recommended across diverse subject, device, and session conditions.
• The proposed adaptive method ensures sustained high performance without requiring a traditional training process.
• This work presents a transferable, adaptive SSVEP-BCI solution that is generalized, plug-and-play, and calibration-free.