Decoding of auditory surprise in adult magnetoencephalography data using Bayesian models

Background: The Bayesian brain framework has been proposed to explain how the brain processes and interprets sensory information. Magnetoencephalography (MEG) and electroencephalography (EEG) are two neuroimaging techniques commonly used with decoding models to study neural responses to auditory, visual and somatosensory stimuli. Our study aims to investigate neural responses to auditory stimuli using MEG data and to determine which temporal components in MEG data are sufficient for decoding surprise based on Bayesian models. Method: MEG data acquired from 18 subjects during an auditory binary oddball task was used. The data were pre-processed, and features were selected from different time windows. Five Bayesian learning models were applied to the experimental task stimuli, and each single trial's surprise value was calculated. The relationship between the extracted features in MEG data and the surprise regressors was investigated using linear regression and 5-fold cross-validation. Results: The results showed that the middle and late components of the MEG evoked potentials were significantly more informative than the early components. The results indicated that the Dirichlet-Categorical model outperformed the other model's decoding performance as demonstrated by higher R-squared values and lower MSE and BIC values. Conclusions: The findings of this study provide evidence for the existence of a neural network that generates surprise in the human brain and highlight the importance of the middle and late components of the MEG evoked potentials for decoding the surprise value of auditory stimuli.