AbstractsComputer Science

Deep Brain Stimulation (DBS) uses surgically-implanted electrodes to stimulate targeted areas of the brain that control movement. DBS provides remarkable therapeutic benefits for many neurodegenerative disorders, including Parkinson’s disease (PD) and Essential Tremor (ET). Existing DBS systems operate open-loop, that is, the stimulation is provided continuously and it does not follow the patient’s disease progression over time. This dissertation proposes a design for next generation automated closed-loop DBS systems, where stimulation is switched on and off depending on when symptoms are predicted to reappear. This thesis work is subdivided into two parts. Part I aims to develop automated tremor prediction algorithms using various machine learning techniques, such as Neural Network (NN), Decision Tree (DT), Tree Bagger (TBAG) and Support Vector Machine (SVM); the algorithm inputs are parameters extracted from surface-Electromyography (sEMG) and Accelerometer (Acc) signals measured at symptomatic extremities. All proposed algorithms achieve high accuracy and sensitivity with low false alarm rate and miss-prediction. The overall best performance is obtained by the DT-based tremor prediction algorithm that achieves 100% sensitivity for all considered ET and PD patients, along with an overall accuracy of 90.0% for ET trials and 85.9% for PD trials. Part II aims to provide a simple yet accurate stochastic model for the neuronal spiking activity in the part of the brain affected by DBS stimulation that could potentially be used to further improve the design of adaptive closed-loop DBS systems. Based on micro-electro recordings obtained during DBS implantation, the parameters of an Ornstein Uhlenbeck Process (OUP) are identified based on the Fortet Integral Equation (FIE) method. The performance of OUP-FIE is compared to other commonly adopted models, such as Brownian Motion (BM), Poisson Process (PP) and the OUP whose parameters are extracted by using the moment method. It is concluded that overall the OUP first passage time distribution is better fitted by the OUP-FIE, for both PD and ET data sets.