Data-driven subtyping of early Parkinson's disease via mutual cross-attention fusion of EEG and dual-task gait features.

Parkinson's disease (PD) exhibits marked clinical heterogeneity, which poses challenges for diagnosis, prognosis, and therapeutic precision, especially for early-stage PD patients. Existing subtyping approaches often rely on subjective clinical scales and single-modality data, which limits their sensitivity in capturing subtle but clinically relevant differences across patients. To reveal clinically meaningful PD subtypes, we propose a data-driven multimodal framework that integrates resting-state electroencephalography (EEG) and dual-task gait features using mutual cross-attention (MCA) fusion. Forty idiopathic early-stage PD patients were enrolled in a prospective study. EEG biomarkers were encoded via a convolutional neural network for the prediction of motor severity (MDS-UPDRS-III), while dual-task gait features were derived to capture subtle motor dysfunctions. The MCA enabled bidirectional attention-guided integration of EEG and gait features, which were then clustered using an unsupervised method. The analysis revealed three distinct subtypes, with dual-task-based fusion providing superior clinical separation. Subtype I was characterized by pronounced motor deficits; Subtype II showed moderate symptoms with relatively preserved quality of life; and Subtype III presented mild motor impairments but exhibited poorer cognitive and psychosocial outcomes. Feature contribution analyses highlighted central beta and theta EEG activity, along with dual-task gait metrics (e.g., stride length during turning), as key drivers of subtype differentiation. Longitudinal follow-up demonstrated subtype-specific rehabilitation responses, with Subtype II showing an insufficient response compared to other subtypes. In conclusion, this study enables digital phenotyping of PD with prognostic implications for personalized rehabilitation strategies and accelerates precision medicine.