An Empirical Analysis of Neurofeedback using PID Control Systems

Neurofeedback systems can be modeled as closed-loop control systems with negative feedback. However, little work to date has investigated the potential of this representation in gaining a better understanding of the actual dynamics of neurofeedback towards explaining subjects' performance. In this paper, we analyze neurofeedback training data through a PID control model. We first show that PID model fitting can produce curves that are qualitatively aligned to the measured BCI signal. Secondly, we examine how brain activity during neurofeedback can be related to common characteristics of control systems. For this, we formalized a pre-existing neurofeedback EEG experiment using a Simulink (R) model that captures both the neural activity and the external algorithm that was utilized to generate the feedback signal. We then used a regression model to fit individual trial data to PID coefficients for the control model. Our results suggest that successful trials tend to be associated to higher average values of K-i, which represents the error-reducing component of the PID controller. It hints that convergence in successful neurofeedback is progressive but complete in approaching the target.