Deep Koopman-operator based model predictive control for closed-loop electrical neurostimulation in epilepsy

Electrical neuromodulation as a palliative treatment has been increasingly applied to epilepsy. However, most of the current neuromodulation implement pre-determined actuation strategies rather than closed-loop neurofeedback. In this paper, rooted in optimal control theory, we propose a novel framework for real-time closed-loop electrical neuromodulation in epilepsy. Our framework combines a deep Koopman-operator based model for seizure prediction in an approximated finite dimensional linear dynamics and the model predictive control (MPC) for designing optimal seizure suppression strategies. We validate our model with synthetic seizure data from Jansen-Rit Model which generates neural dynamics in a single cortical column and two distant cortical columns. The results demonstrate that the deep Koopman-operator based model has great capabilities to map the nonlinear neural dynamics into finite dimensional linear dynamics, which is suitable for real-time seizure prediction and naturally compatible with the optimal-based linear MPC design for seizure suppression. Our framework opens a new window for the development and implementation of robust real-time closed-loop electrical neuromodulation in epileptic seizure suppression and sheds light on understanding the neurodynamics and feedback control policies.