Predicting Motor Trajectories and Mapping Progression Subtypes in Parkinsons Disease via Structure-Function Neural Field Encoding and Multi-View Representation

Authors

Abstract

Parkinsons disease (PD) is characterized by substantial heterogeneity in progression patterns, posing major challenges for individualized prognosis and clinical management. This study presents Structure-Function Neural Field Alignment and Multi-View Distillation (SFNFA-MVD), a novel deep learning framework that integrates structural, functional, and diffusion MRI to predict motor trajectories and identify progression subtypes of PD. SFNFA-MVD is built around two complementary components: a neural field encoding backbone that maps structural, functional, and diffusion MRI into a continuous structure-function field for fine-grained multimodal alignment, and a multi-view distillation module that leverages transformer-based cross-modal fusion to capture long-range dependencies across modalities and timepoints and generate trajectory-aware clinical predictions. Applied to a longitudinal cohort of 268 PD patients from the PPMI dataset, SFNFA-MVD achieved state-of-the-art accuracy in predicting MDS-UPDRS II and III scores (MAE = 2.167 and 2.164; r = 0.613 and 0.695, p < 0.001), outperforming all benchmarked unimodal and multimodal models. Moreover, unsupervised clustering of fused representations revealed three progression subtypes, slow progressing, moderate progressing, and fast progressing, aligned with distinct alterations in the sensorimotor, dorsal attention, and salience networks, respectively. Importantly, functional network density within each subtype correlated significantly with longitudinal motor decline (r = 0.53-0.57, p < 0.01). These findings demonstrate that SFNFA-MVD enables interpretable, individualized modeling of PD progression, offering a powerful framework for precision prognosis and network-level stratification. Author summaryPD causes gradual loss of movement control and affects people differently, making it hard for doctors to predict how symptoms will change over time. This study developed a new artificial intelligence tool, SFNFA-MVD, that learns from multiple types of brain scans to better forecast individual motor progression. Using structural, functional, and diffusion MRI from 268 people in a long-term Parkinsons cohort, the model combines information across brain structure and brain activity to predict future motor symptom severity. We found that SFNFA-MVD predicted motor scores more accurately than existing single-scan or simple fusion approaches. Beyond prediction, it also grouped patients into three progression patterns--slow, moderate, and fast--each linked to distinct changes in brain networks involved in movement, attention, and salience processing. Importantly, the degree of network disruption within each group tracked with worsening motor symptoms over time. By connecting imaging-based brain changes with personal symptom trajectories, this approach may help clinicians monitor PD more precisely, identify higher-risk patients earlier, and support more personalized disease management.

Tags

Topics