Dual-Branch Fusion Transformer Networks for Remote Parkinson's Disease Screening via Hand-Drawn Graphics

Parkinson's disease (PD) is one of the most common neurodegenerative disorders in the aging population, which is commonly associated with motor abnormalities that can be reflected in hand-drawn geometric patterns. However, early detection remains challenging due to the lack of specific screening or tests, especially in rural areas with limited medical resources, leading to delays in medical intervention. To address these critical limitations, this paper proposes FusionNet, a dual-branch vision Transformer framework that combines spatial and frequency-domain information for handwriting-based PD screening. The spatial branch learns geometric and stroke-related representations directly from grayscale drawings, whereas the frequency branch processes log-scaled amplitude spectra obtained by two-dimensional fast Fourier transform. Features from the two branches are integrated using bidirectional cross-attention (BCA), allowing spatial and spectral representations to condition one another before classification. A staged training strategy was adopted in which the individual branches were trained separately before optimization of the fusion components. Experimental evaluations on the publicly available NewHandPD dataset were conducted across the three drawing tasks: hand-drawn circle, meander, and spiral images from 66 participants (31 individuals with PD and 35 healthy controls). Performance was evaluated using five-fold Group K-Fold cross-validation, demonstrating that FusionNet achieves the highest accuracy of 91.68% on the meander task and maintains >85% sensitivity with >88% specificity across all tasks. Because the study used a small single dataset, participant-level separation was not guaranteed for all tasks, and no independent external validation was performed, the reported results should be interpreted as preliminary and dataset-specific. Analysis of model behavior further indicates that spatial-domain predictions could be affected by paper texture, whereas the complementary frequency representation reduces this source of bias. Finally, the model was incorporated into a LINE-based prototype to demonstrate the technical feasibility of remote image submission and automated analysis. This implementation remains a proof of concept and requires external, prospective clinical validation before use in clinical decision-making.