Enhancing SSSEP-Based BCI Performance with Adaptive GMM and Personalized Channel Selection

Brain-computer interfaces (BCIs) provide a direct communication pathway between the human brain and external devices. Tactile BCIs have emerged as a promising approach for visually impaired users and for developing multisensory hybrid BCIs. Among tactile BCIs, steady-state somatosensory evoked potential (SSSEP)-based BCI is a representative paradigm, yet it faces challenges including relatively low classification accuracy and significant inter-subject variability. This paper proposes a personalized SSSEP-based BCI framework that integrated individual resonance-like frequency identification, Common Spatial Pattern (CSP)-based channel selection, and an adaptive Gaussian Mixture Model (GMM) classifier. The framework first identified individual resonance-like frequencies under continuous tactile stimulation through a frequency-scanning procedure, then applied CSP to extract subject-specific spatial patterns of SSSEP features during a selective attention task, enabling personalized channel selection and low-dimensional feature representation. These personalized features were modeled using a GMM classifier, whose parameters were subsequently updated with online data to further improve decoding performance. Experiments with eight participants performing selective tactile attention tasks on their left and right index fingertips demonstrated an average offline classification accuracy of 83.86%±5.69%, with six participants exceeding 80% and one exceeding 90%. Online experiments involving four participants showed a progressive increase in classification accuracy across 16 sessions, reaching performance levels comparable to offline results. These findings indicate that the proposed framework effectively enhanced classification accuracy and robustness while reducing inter-subject variability in SSSEP-based BCIs.