Prediction of mild Parkinsonism revealed by neural oscillatory changes and machine learning.

Neural oscillatory changes within and across different frequency bands are thought to underlie motor dysfunction in Parkinson's disease (PD) and may serve as PD biomarkers. Here, we used oscillatory signals derived from chronically-implanted cortical and subcortical electrode arrays as features to train machine-learning algorithms to classify the clinical states in a nonhuman primate model. LFP data were collected over several months from a 12-channel subdural ECoG grid and a 6-channel custom array implanted in the subthalamic nucleus (STN). Relative to the naive state, PD-state showed elevated M1 and STN power in the beta, high-gamma, and HFO bands, but decreased power in the delta band. Theta power decreased in STN but not M1. In the PD state there was elevated beta-HFO phase-amplitude coupling (PAC) in the STN. We applied support vector machines with Radial Basis Function (SVM-RBF) kernel, and k-nearest neighbors (KNN) classifiers trained by features related to power and PAC changes to discriminate between the naive and PD states. Our results show that the most predictive feature of Parkinsonism in the STN was high-beta (~86% accuracy), and HFO in M1 (~98% accuracy). A feature fusion approach outperformed every individual feature in the STN (~96% accuracy). Overall, our data demonstrate the ability to use different frequency band powers in classifying clinical state, including a potential added benefit of feature fusion approaches, even during a relatively mild stage of the disease. This type of feature-tailored approach (using single or multiple features) may contribute to further optimizing patient-specific closed-loop or adaptive DBS approaches.