Volitional modification of brain activity in adolescents with Autism Spectrum Disorder: A Bayesian analysis of Slow Cortical Potential neurofeedback.

Autism spectrum disorder is (ASD) characterized by a persisting triad of impairments of social interaction, language as well as inflexible, stereotyped and ritualistic behaviors. Increasingly, scientific evidence suggests a neurobiological basis of these emotional, social and cognitive deficits in individuals with ASD. The aim of this randomized controlled brain self-regulation intervention study was to investigate whether the core symptomatology of ASD could be reduced via an electroencephalography (EEG) based brain self-regulation training of Slow Cortical Potentials (SCP). 41 male adolescents with ASD were recruited and allocated to a) an experimental group undergoing 24 sessions of EEG-based brain training (n1 = 21), or to b) an active control group undergoing conventional treatment (n2 = 20), that is, clinical counseling during a 3-months intervention period. We employed real-time neurofeedback training recorded from a fronto-central electrode intended to enable participants to volitionally regulate their brain activity. Core autistic symptomatology was measured at six time points during the intervention and analyzed with Bayesian multilevel approach to characterize changes in core symptomatology. Additional Bayesian models were formulated to describe the neural dynamics of the training process as indexed by SCP (time-domain) and power density (PSD, frequency-domain) measures. The analysis revealed a substantial improvement in the core symptomatology of ASD in the experimental group (reduction of 21.38 points on the Social Responsiveness Scale, SD = 5.29), which was slightly superior to that observed in the control group (evidence Ratio = 5.79). Changes in SCP manifested themselves as different trajectories depending on the different feedback conditions and tasks. Further, the model of PSD revealed a continuous decrease in delta power, parallel to an increase in alpha power. Most notably, a non-linear (quadratic) model turned out to be better at predicting the data than a linear model across all analyses. Taken together, our analyses suggest that behavioral and neural processes of change related to neurofeedback training are complex and non-linear. Moreover, they have implications for the design of future trials and training protocols.