Patients with Parkinson’s disease exhibit differential local field potential (LFP) activities under different deep brain stimulation (DBS) conditions. To achieve a balance between the discriminative performance and computational complexity for the identification of DBS stimulation-related neural states, this paper proposes a feature-division ensemble learning framework. First, multi-objective feature selection was conducted based on stability, discriminability and parsimony to derive a compact feature subset from time-domain, frequency-domain and nonlinear features. Then, the three categories of features were assigned to their respective advantageous base classifiers and fused via hard voting. Window-level 10-fold cross-validation results on the public STN-LFP dataset from the University of Oxford demonstrated that the proposed model reached an accuracy of 97.18%, outperforming the support vector machine (88.85%) and eXtreme Gradient Boosting (92.84%). SHAP analysis revealed that β-related frequency-domain features and partial nonlinear features had relatively high contribution. The feature division strategy reduced the cumulative input dimension of the three base classifiers from 96 to 32 and eliminated redundant feature processing. This method can provide a reference for analyzing DBS stimulation-related LFP patterns and further research on closed-loop feedback biomarkers.
