Deep Brain Recordings in Parkinson's Disease: Processing, Analysis and Fusion with Anatomical Models

Abstract

This thesis presents several novel techniques and tools for automatic classi cation and analysis of highly detailed invasive recordings of the brain activity in patients with Parkinson's disease (PD). By utilizing machine learning concepts, we approach three of the principal questions, central to modern treatment and understanding of the PD: i) What information about patient's state can be derived from recorded brain activity? By identifying patterns characteristic for tremor onset in signals recorded through deep brain stimulation electrodes, we show that an adaptive system, modifying treatment parameters to match current state of its bearer, is feasible. ii) How to obtain trustworthy answers to scienti c questions from noisy microelectrode activity recordings? We show that undesirable noise is highly prevalent in intraoperative microelectrode recordings and provide the sigInspect: a GUI tool for annotation of microelectrode signals. The tool includes a set of well-performing classi ers for automatic artifact identi cation, validated on an extensive multi-center database of manually labeled data. iii) Where exactly in the target nucleus were the signals recorded? This question is vital for appropriate stimulation electrode placement as well as for better understanding of possible side e ects. We propose a novel probabilistic model for tting a 3D anatomical atlas of the subthalamic nucleus based solely on the recorded electrophysiological activity and show that such approach may lead to more accurate localization of recording sites during and after the surgery.