Classical conditioning of the BOLD signal as a paradigm for basic BCI communication in Alzheimer patients

Abstract

Background Alzheimer patients express their need of social interaction even when their communication abilities are highly impaired. (Mayhew et al., 2001). Brain-computer interfaces (BCI), already used with severely paralyzed patients, may be adapted for communicating with Alzheimer patients by shifting the paradigm from instrumental-operant learning to classical conditioning (Birbaumer, 2006), e.g. by associating “yes” thinking to a positive emotion and “no” thinking to a negative emotion. We designed an fMRI-based BCI setting aimed at conditioning subjects to associate positive and negative emotional stimuli with respectively congruent and incongruent word pairs, in view of a BCI application for basic yes/no communication. Methods fMRI was performed on 6 healthy subjects during a classical conditioning session, comprising the phases of habituation, acquisition and extinction. The unconditioned stimuli consisted of a positive (baby laughter) and negative (scream) emotional sound. The conditioned stimuli, presented aurally, were congruent (e.g. “animal-elephant”) and incongruent (e.g. “animal-Germany”) word pairs. During the conditioning acquisition phase, congruent word pairs were associated to the baby laughter and incongruent word pairs were associated to the scream. A linear Support Vector Machine was implemented to classify the BOLD signal corresponding to congruent and incongruent word pairs. To investigate the relative importance of distinct brain areas in decoding different brain states, feature vectors from the frontal cortex were also used as input to build a separate classifier. Results Participants rated in the Self Assessment Manikin (SAM) the scream and the laughter as having, respectively, negative and positive valence. Moreover, the scream was associated to higher arousal compared to the baby laughter. Classification of the BOLD signal as a response to the congruent and incongruent word pairs immediately followed by the emotional unconditioned stimuli showed above chance level performance (57-64%) in one subject, while the performance was chance level on the remaining subjects (44-56%). Conclusions We presented a possible paradigm for basic yes/no communication, based on the conditioning of BOLD signal by the repeated association of emotional and semantic stimuli. The performance of the classifier by feature selection needs to be improved before being tested with Alzheimer patients. Future developments also comprise an online implementation of the system.