Research On Neural Computational Model-Driven Closed-Loop Deep Brain Stimulation

Deep Brain Stimulation(DBS)is an effective treatment for Parkinson’s disease.It improves brain function by applying high-frequency electrical impulses to regulate neural activity in focal area of Parkinson’s disease.The key to improving the effect of DBS is the understanding of the neural circuit in the focal area of Parkinson’s disease.Establishing a neural computational model of Parkinsonian state is essential for investigating the biophysical mechanisms of Parkinson’s disease,analyzing the mechanism of DBS actions,and further improving the effects of DBS therapies.In clinical practice,open-loop DBS parameters,including frequency,amplitude and pulse width,are selected according to the doctors’ experience,which greatly limits the effect of DBS.Therefore,it is of great significance to explore efficient strategies for the optimization of DBS parameters.Moreover,designing a closed-loop DBS,which could adaptively adjust DBS parameters according to real-time feedback of biomarkers,is one of the effective approaches to achieving precise control of the Parkinsonian state.In this thesis,neural computational model-driven closed-loop DBS is proposed.Biomarkers representing the Parkinsonian state are extracted from the constructed neural computing model of the Parkinsonian focus area.The reinforcement learning algorithm is introduced for the optimization of DBS parameters and the adaptive algorithms are exploited for achieving closed-loop DBS.Major contributions of the thesis are listed below:(1)According to the neuronal dynamics,synaptic dynamics and network topology structure,a neural computational model of the Cortex-Basal ganglia-Thalamus Circuit(CBTC)is established to characterize the abnormal neural activities in the focal area of Parkinson’s disease.The Parkinsonian state that induced by the absence of dopamine neurotransmitters is simulated,and the differences in neural activity between normal state and Parkinson’s state are quantified.The simulation results illustrate that the enhancement of Beta band oscillations in the local field potential signals significantly relates to the Parkinson’s disease.(2)Inspired by reverse engineering,a CBTC-based recurrent neural network is built according to abnormal oscillations in the Parkinsonian state,through the targetbased first order reduced and controlled error algorithm.The rhythmic oscillations under the Parkinsonian state are reconstructed,the effect of DBS is simulated,and the effectiveness of high-frequency DBS is further proved.The variational autoencoder is used to extract the low-dimensional manifold trajectories of the recurrent neural network,by which,the differences between diverse rhythmic oscillations in the Parkinsonian state are revealed from the perspective of dynamics,and the internal dynamic transition mechanisms under DBS are characterized.(3)A reinforcement learning-based algorithm is introduced for the optimization DBS parameters.The neural computational model of the CBTC is adopted as the interactive environment,and the reward function is constructed to suppress Beta band oscillations and the reduce stimulus energy consumption.Deep Q Network algorithm and Actor-Critic algorithm are exploited to train agents,for optimizing DBS parameters in discrete space and continuous space,respectively.The results show that the reinforcement learning-based DBS parameter optimization algorithm can effectively guide the selection of DBS parameters.(4)A closed-loop DBS strategy that driven by the CBTC neural computational model is proposed,to achieve precise regulation of the Parkinsonian state.To improve the instantaneous effect of DBS,a minimum variance closed-loop DBS strategy is designed based on the local field potential signals.To improve the long-term effect of DBS,a neural network closed-loop DBS is designed based on Beta band power.The results prove that the proposed closed-loop DBS strategy can adapt to the changes of personalized parameters,ultimately achieving effective suppression of Parkinson’s state.In summary,a neural computational model of the Parkinsonian focal area is constructed to describe the abnormal rhythm oscillations of the Parkinsonian state.The proposed parameter optimization algorithm and closed-loop DBS strategy provide new insights into intelligent neural regulation therapy.