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

Wang, Deyu; Shi, Yu; Pang, Jun; Zhu, Xiaodong; Meng, Lin; Ming, Dong · 2026 · DOAJ

DOI: 10.1038/s41531-026-01258-2

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the clinical heterogeneity of early-stage Parkinson’s disease (PD), which complicates diagnosis and personalized treatment. Existing subtyping methods often rely on subjective clinical scales or single-modality data, failing to capture subtle neurophysiological and motor differences. To overcome this, the authors developed a data-driven multimodal framework integrating resting-state electroencephalography (EEG) and dual-task gait features using mutual cross-attention (MCA) fusion. The goal was to identify robust, clinically meaningful PD subtypes and evaluate their longitudinal responsiveness to rehabilitation. The study enrolled 40 idiopathic early-stage PD patients. Data acquisition included 32-channel resting-state EEG recordings and motor assessments involving both single-task and dual-task walking (walking while performing serial-3 subtraction). EEG power spectral density features were encoded via a convolutional neural network to predict motor severity, while 24 gait parameters were extracted from inertial measurement units. The MCA mechanism fused these heterogeneous modalities by dynamically weighting cross-modal dependencies, followed by unsupervised k-means clustering to identify subtypes. The analysis revealed three distinct subtypes, with dual-task fusion providing superior clinical separation compared to single-task data. Subtype I (n=8) exhibited pronounced motor deficits, including rigidity and impaired dexterity. Subtype II (n=15) showed moderate symptoms but relatively preserved quality of life and elevated frontal/parietal beta EEG activity. Subtype III (n=17) presented mild motor impairments but poorer cognitive and psychosocial outcomes. Feature contribution analysis identified central beta and theta EEG activity, along with dual-task gait metrics such as stride length during turning and hip range of motion, as key drivers of differentiation. Longitudinal follow-up after a 3-month rehabilitation program showed subtype-specific responses: Subtype I had the largest motor improvement, while Subtype III showed significant correlations between baseline pelvic kinematics and therapeutic gains, unlike Subtypes I and II. The findings demonstrate that integrating EEG and dual-task gait data via attention-based fusion enables precise digital phenotyping of PD. The identified subtypes have distinct neurobehavioral signatures and differential rehabilitation responses, suggesting that baseline multimodal features can predict treatment outcomes. This approach offers a pathway for precision medicine, allowing for subtype-specific interventions and improved clinical management in early-stage PD.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms