Bayesian multi-subject factor analysis to predict microsleeps from EEG power spectral features

Prediction of an imminent microsleep has the potential to save lives and prevent catastrophic accidents. A microsleep is a brief episode of unintentional unconsciousness and, hence, loss of responsiveness. In this study, prediction of imminent microsleeps using EEG data from 8 subjects was examined. A novel Bayesian algorithm was proposed to identify common components of pre-microsleep activity in the EEG in all subjects and predict microsleeps 0.25 s ahead. To avoid overfitting, this model incorporates sparsity-promoting priors to automatically find the minimum number of components. Due to intractability of full Bayesian treatment, variational Bayesian was integrated to approximate posterior probabilities. To predict microsleeps, EEG log-power spectral features were extracted from a 5-s window. Bayesian multi-subject factor analysis was used to extract common microsleep patterns and transform all features into lower-dimension common-space features. Discrimination between responsive and microsleep instances was done with a single linear discriminant analysis (LDA) classifier. Performance of the proposed method was evaluated using leave-one-subject-out cross-validation. Our prediction system achieved moderate AUC
_ROC
and GM of 0.90 and 0.80, respectively, but with a relatively low precision of 0.29.