Patient-Optimized Temporally-Adaptive Neurostimulation for Epilepsy Using Deep-EDMD Koopman MPC

This thesis presents the design and implementation of a closed-loop neurostimulator controller for delivering patient-optimized, temporally-adaptive stimulation pulses to control epileptic seizures. The method employs a predictive model and a Model Predictive Controller (MPC) to optimize stimulation. The predictive model, based on a deep learning extended dynamic mode decomposition algorithm, approximates the Koopman operator, capturing a patient's brain dynamics and forecasting stimulation efficacy. The MPC uses these predictions to converge rapidly to an optimal set of stimulation parameters. The system is capable of adapting to changes in brain dynamics over time, ensuring continuous optimization. Both the predictive model and MPC are implemented in software and on hardware (FPGA and ASIC synthesis), achieving a power consumption of 97.09 μW.

Additionally, we present a framework incorporating a neural mass model (NMM) fine-tuned to patient pre-recorded data. The NMM generates synthetic intracranial electroencephalography (iEEG) highly correlated with real iEEG during normal and seizure periods. This framework is used to test and validate other patient-optimized, temporally-adaptive stimulation approaches. Our temporally-adaptive stimulation optimization for seizure control was validated using this framework.