465 Classifying At-Home Movement States Using Cortical-Pallidal Neural Activity in Parkinson's Disease

INTRODUCTION: The development of adaptive deep brain stimulation (aDBS) for Parkinson’s disease (PD) to automatically change stimulation from conventional to ‘gait-optimized’ frequency settings when patients walk is challenged by the lack of neural biomarkers of movement states in naturalistic environments. METHODS: A bi-directional neurostimulator (Summit RC+S, Medtronic) was used to record unilateral activity from the globus pallidus (GP), premotor cortex (PM), and primary motor (M1) cortex in two PD patients. Wearable ankle accelerometers (Rover Health) were used to record at-home kinematic data. Rover and RC+S signals were synchronized, and spectral analysis of 10-second epochs of continuous walking and non-walking was performed. Linear discriminant analysis (LDA) models were trained to classify movement states using average power within all possible frequency bands from 1 to 50 Hz during each epoch. The most important features determined by random forests (RF) were used to create ‘system-constrained’ LDA models meeting the specifications of the RC+S on-board classifier. RESULTS: Over 16 hours of data were analyzed (subject 1: 10.75 hours, subject 2: 5.7 hours). Spectral profiles of movement state differed between the two subjects. While one exhibited broadband differences across multiple frequency ranges in cortical-pallidal signals between states, the second patient displayed more focused decreases in M1 beta band (13–30 Hz) power during walking periods. Models achieved areas under the curve (AUC) greater than 0.80 in both subjects when trained with all potential frequency bands. ‘System-constrained’ classifiers performed with AUC greater than 0.70 for both individuals. CONCLUSIONS: Our findings support the hypothesis that motor cortex and basal ganglia oscillations are modulated by movement state. These novel methods offer a pipeline for identifying patient-specific movement state biomarkers using long-term naturalistic recordings of neural and kinematic activity.