A multimodal mutual information‐guided feature selection framework for predicting rehabilitation response in Parkinson's disease with postural instability and gait disorder

Abstract Postural instability and gait disorder (PIGD) subtype of Parkinson's disease (PD) is marked by heterogeneous motor and cognitive impairments, making rehabilitation response difficult to predict. Identifying robust multimodal predictors is essential for precision rehabilitation. This study aimed to identify key multimodal features associated with response to motor‐cognitive interactive rehabilitation and to develop a generalizable prediction framework. Twenty‐one PD patients with PIGD completed a motor‐cognitive interactive rehabilitation program. Multimodal data, including demographics, clinical scales, gait parameters, magnetic resonance imaging (MRI), and EEG, were collected across 14 feature modalities. A multimodal sequential forward selection framework based on mutual information (MSFSF‐MI) was proposed where predictive stability of selected feature sets was assessed across five machine learning models (support vector machine, RBF, random forest, stochastic gradient boosting, and XGB). Multimodal feature subsets derived by the proposed framework consistently outperformed unimodal models across classifiers. Cross‐model analyses highlighted functional connectivity, cortical thickness, low‐frequency power spectral density, and phase–amplitude coupling as reproducible predictors, forming a key feature set mainly from MRI and EEG domains. This study identified predictive and potentially robust multimodal neural features of PD rehabilitation response. The introduced nested prediction framework demonstrates strong potential for future generalization, providing a methodological foundation for personalized neurorehabilitation strategies.