Classification and characterization of neural activity during bradykinetic and freezing of upper limb movements

Abstract

Bradykinesia and freezing of the upper limb (FOUL) movements are debilitating symptoms of Parkinson’s disease (PD) that occur during rapid alternating movements (RAMs). When attempting to perform high-rate RAMs, movements are typically characterized by bradykinesia (maintaining amplitude by slowing velocity) or episodes of FOUL (hesitations, freezing, or festination - hastening with small amplitudes). Currently, little is known about the neural signatures of basal ganglia oscillations associated with bradykinesia and FOUL behaviors. In this study, ten people with PD with externalized deep brain stimulation (DBS) leads (directional 1-3-3-1 configuration) implanted in the globus pallidus internus (GPi) performed a RAMs task while OFF medication and OFF stimulation. Participants performed epochs (8-12 s) of forearm pronation-supination movements (±15 deg.) that were cued by an acoustic stimulus alternating between slow (1Hz) and fast (2Hz) rates. For each of the 226 fast epochs, power spectra of bipolar local field potentials (LFPs) were calculated for all vertically adjacent DBS contact pairs using bipolar ring pairs (containing dorsal-ventral information) or directional pairs (containing circumferential information). Each epoch was visually classified from the forearm angular position traces as normal (met criteria for movement amplitude and rate), bradykinetic (low velocity at target amplitude), or FOUL (hesitations, freezing, festination). Five machine learning models were trained to distinguish between the three movement classes, and SHapley Additive exPlanations (SHAP) analysis was used to identify the frequency bands that contributed most towards classification, for each of the LFP pair sets (ring versus directional pairs). There was no significant difference found in model performance between features using bipolar ring pairs or directional pairs. Across both pair types, the non-linear Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP) models significantly outperformed the linear Gaussian Naïve Bayes (GNB) and Linear Discriminant Analysis (LDA) models. For the RF, SVM and MLP models, frequencies within β-band (13-33Hz) and θ-band (4-8Hz) were consistently ranked among the most important features, however when comparing the power in individual frequency bands between movement classes, there were no statistical differences shown between any of the top features. This suggests that there are non-linear relationships between frequency bands that distinguish between the movement classes, and that bradykinesia and FOUL may be distinct neural phenomena.