UC San Diego

Brain-computer Interface(BCI) provides a direct communication pathway between the targeted brain region by recording its evoked neural signals and control systems that interact with neural interfaces. Our work is aiming to find an alternative brain-computer interface (BCI) control scheme to decode affective states, for assessing and interpreting changes in the user state while evaluating goal-oriented control schemes, thus improving the efficiency, usability, and accuracy of the BCI. To build such a BCI system, successfully decoding the affective state is an essential step to take. There are several neural studies with non-invasive recording techniques that shed insight on the complex and subtle relationships between affective state and neural response. However, it is challenging to capture the high-resolution spatio-temporal patterns in the neural response with non-invasive recording techniques. The precise spatio-temporal pattern in cortical depth structures is important to characterize complex affective states at different time scales.Addressing the complexity of neural representation, we bring an acoustic perception experiment into the Epilepsy Monitoring Unit (EMU) and obtain intracranial neural recordings from subjects listening to the natural acoustic stimuli that spans different dimensions of the affective states. In this work, we present several decoder models with different levels of regularization (ordered by descending model complexity: quadratic discriminative analysis (QDA), regularized QDA with class-specific Gaussian process factor analysis (GPFA) and regularized QDA with class-invariant GPFA). Through an iterative process of feature selection and model simplification, we identify few models that emphasize the most informative feature for decoding effect.