Parkinson's disease severity is encoded by non-linear oscillatory interactions across cortical-subcortical networks.

Parkinson's Disease (PD) is characterized by motor symptoms linked to dysregulated neural activity within basal ganglia-thalamocortical (BGTC) circuits. While clinical assessments such as the Unified Parkinson's Disease Rating Scale (UPDRS) effectively measure symptom severity, they are not designed to provide insights into underlying neuropathophysiology. We hypothesize that PD severity is encoded by nonlinear interactions across BGTC neurophysiological biomarkers. We recorded simultaneous preiriolandic electrocorticography and globus pallidus internus (GPi) local field potentials in 47 patients undergoing deep brain stimulation (DBS) implantation. Using 110 spectral features, we compared linear regression against a non-linear ensemble of bagged decision trees to predict preoperative UPDRS scores (off medicine). Linear regression failed to robustly predict severity. Conversely, the non-linear ensemble model successfully predicted motor scores (cross-validation rho = 0.317, p = 0.018). Predictive power emerged from a distributed pattern of biomarkers, including alpha, low-beta, and gamma burst metrics and coherence, rather than a single dominant feature, indicating that PD pathophysiology is encoded by collective network states rather than isolated oscillations. Consequently, accurate, subject-specific severity estimation requires multivariate, non-linear modeling, providing deeper insight into the stable, baseline pathological network states upon which dynamic symptom fluctuations occur.