CRCNS: US-GERMAN RESEARCH PROPOSAL: Deep Neural Network Approaches for Closed- Loop Deep Brain Stimulation Using Cortical and Subcortical Sensing Principal Investigators: R. Mark Richardson, MD, PhD, Department ofNeurologicaJ Surgery and Robert S. Turner, PhD, Department ofNeurobiology, University ofPittsburgh; Wolf-Julian Neumann, MD, and Andrea A. Kiihn, MD, Department ofNeurology, Charite-Universitatsmedizin Berlin. Co-Investigators: Benjamin Blankertz, PhD, Department of Computer Science, Technische Universitat Berlin; Tom Mitchell, PhD, Machine Learning Department, Carnegie Mellon University. PROJECT DESCRIPTION 1 Introduction and Objectives Deep brain stimulation (DBS) represents one of the major clinical breakthroughs in the age of translational neuroscience.1 In 1987, Benabid and colleagues demonstrated that high frequency stimulation can mimic the effects of ablative neurosurgery in Parkinson's disease (PD)2'3, while offering two key advantages to previous procedures: adjustability and reversibility. This project will employ artificial intelligence strategies, which have allowed for unprecedented innovations in translational, personalized, high-definition medicine4, to further elevate the therapeutic potential of DBS. By developing a computational framework for decoding behavioral and disease states from combined subthalamic and cortical population recordings, this work will inform future adaptive stimulation paradigms for PD and other movement disorders. The central aim is to develop a computational framework for deep learning- based multi-feature decoding of behavioral and disease states from electrocorticography (ECoG), in order to advance the evolution of aDBS. Proposed Research. The concurrent use ofresearch ECoG during DBS surgery recently has enabled basic neuroscience investigation ofhuman co11ical-subcortical network dynamics. The overall goal ofthis project is to establish intelligent algorithms to identify physiological and pathophysiological states in ECoG data that predict epochs during which stimulation would facilitate movement or reduce symptoms. Given that ECoG electrode strips can be implanted safely during DBS surge1y56, the use of ECoG for aDBS is hypothesized to have several advantages: 1. compatibility with any DBS electrode design; 2. optimal signal to noise ratio, 3. customization of electrode implantation location, based on known network connectivity of the target nucleus; 4. potential implantation of multiple remote sensors in different brain regions. Several barriers, however, exi'st for advancing this type of strategy. Namely, neither instantaneous biomarkers, evidence for optimal recording locations, adaptive control algorithms, nor previous reports on the use of ECoG-based, adaptive DBS have been described in PD patients. These obstacles will be overcome through achieving the following: Objective #1: Identify cortical and subthalamic biomarkers that can distinguish activity related to physiological motor behaviorfrom that related to pathological symptoms. The goal of objective 1 is to distinguish oscillatory features in population activity based on the temporal, spectral and spatial specificity of their synchronization across the corticosubthalamic axis, e.g. characteristic evolution of beta activity with bradykinesia vs. rhythmic high frequency activity synchronized during tremor. Features that reflect instantaneous motor behavior and parkinsonian symptom severity will be characterized through spectral decomposition mapped to 3D reconstructions of recording locations. Objective #2: Determi11e tlte optimal computational approaclt fol' real-time ,letection and prediction of physiological andpatliophysiological behaviors. The goal of objective 2 is to use results from the first objective to build a computational framework for deep learning of physiological and pathophysiological states from cortical time-series data based on hierarchical recurrent artificial neural networks. Joint data collection at the two high-volume DBS centers 37
