Parkinson’s Disease State Recognition from LFP: An Ensemble Model with Feature Optimization

Parkinson's disease (PD) is a debilitating neurodegenerative disorder.While deep brain stimulation (DBS) alleviates symptoms, advancing closed-loop DBS requires real-time, accurate decoding of patient states from local field potential signals.Traditional machine learning models face a trade-off between accuracy and computational efficiency.This study proposes a feature-optimized ensemble framework.It first employs multi-objective optimization for feature selection to identify an optimal subset, reducing dimensionality while preserving discriminative information.Based on this, a feature-partitioned ensemble model under a hard voting framework is constructed, allocating distinct feature types to specialized base classifiers.Evaluation on a public PD dataset confirms the model's superiority, achieving a leading accuracy of 97.18%.The results show that the ensemble leverages diverse algorithms via hard voting, while its feature-partitioning design avoids the computational burden of processing full high-dimensional feature sets.SHAP interpretability analysis identified key decision-driving features, reaffirming the critical role of beta-band activity in PD state discrimination.