Precision prediction of motor fluctuations in advanced Parkinson’s disease via multiregional neural sensing

Abstract Precise prediction of motor fluctuations is critical for the management of advanced Parkinson’s disease (PD). However, current single-region biomarkers show poor performance largely due to their mismatch with the network-level pathophysiology. This study presents a translatable multi-regional sensing framework for simultaneous recording from eight cortico-basal ganglia-thalamo-cortical (CBGT) circuit hubs in rodent models. Application of a hierarchical loss-optimized multilayer perceptron (H-Lo-MLP) tailored for open-set clinical scenarios demonstrates that distributed CBGT oscillatory patterns substantially outperform single-region biomarkers. The H-Lo-MLP achieves nearly 90% prediction accuracy, far exceeding the ~50% of threshold-based approaches, with robust generalizability across disease progression and dopaminergic modulation. Systematic evaluation of all 255 region combinations identifies primary motor cortex (M1) and dorsal striatum as pivotal discriminative hubs and reveals a parsimonious four-region configuration that retains near-optimal predictive performance while minimizing surgical invasiveness. Cross-species proof-of-concept in a single non-human primate recapitulates the conserved centrality of M1 and demonstrates the feasibility of simultaneous multi-region monitoring via a single linear electrode. These findings establish network-level sensing as a clinically viable paradigm for precise motor fluctuation prediction, overcoming limitations of constrained adaptive deep brain stimulation. By validating a surgically tractable, cross-species network-sensing approach, this work bridges the gap between circuit-level pathophysiology and clinical translation.