Slow Cortical Potential BCI Classification Using Sparse Variational Bayesian Logistic Regression with Automatic Relevance Determination

Detecting P300 slow-cortical ERPs poses a considerable challenge in signal processing due to the complex and non-stationary characteristics of a single-trial EEG signal. EEG-based neurofeedback training is a possible strategy to improve the social abilities in Autism-Spectrum Disorder (ASD) subjects. This paper presents a BCI P300 ERPs based protocol optimization used for the enhancement of joint-attention skills in ASD subjects, using a robust logistic regression with Automatic Relevance Determination based on full Variational Bayesian inference (VB-ARD). The performance of the proposed approach was investigated utilizing the IFMBE 2019 Scientific Challenge Competition dataset, which consisted of 15 ASD subjects who underwent a total of 7 BCI sessions spread over 4 months. The results showed that the proposed VB-ARD approach eliminates irrelevant channels and features effectively, producing a robust sparse model with 81.5 ± 12.0% accuracy in relatively short modeling computational time 19.3 ± 1.4 s, and it outperforms the standard regularized logistic regression in terms of accuracy and speed needed to produce the BCI model. This paper demonstrated the effectiveness of the probabilistic approach using Bayesian inference for the production of a robust BCI model. Considering the good classification accuracy over sessions and fast modeling time the proposed method could be a useful tool used for the BCI based protocol for the improvement of joint-attention ability in ASD subjects.