Abstract
INTRODUCTION: Subtle PD motor abnormalities can be underappreciated in examination but are overtly present in handwriting. Spiral, meander, and wave drawings are noninvasive, low-cost methods for capturing PD motor signatures and are easily collected on a tablet. Deep learning is a natural framework to decode such features, but most handwriting-based PD classifiers have not been thoroughly optimized or tested for generalization, leaving their clinical utility unverified. METHODS: PD-MGMA-DSCNN is a multiscale gated multi-head attention depthwise separable convolutional network operating on a unified RGB image of spiral, meander, and wave tasks. A PD-specific Bayesian-genetic optimization scheme (PD-BayGA) was used to tune a suite of key architectural and training hyperparameters. Model decisions were interpreted using SHAP-based superpixel attribution maps to link discriminative regions to previously described Parkinsonian handwriting abnormalities. RESULTS: On the PaHaW dataset, PD-MGMA-DSCNN achieved 98.23% accuracy on the held-out internal test set, outperforming CNN, RNN, 3D CNN, and ensemble baselines reported in the literature. Using the finalized configuration developed on PaHaW, external evaluation on the independent HandPD dataset yielded 99.42% accuracy, together with high ROC-AUC and precision-recall performance under clinically motivated rule-out and rule-in thresholds. These findings support the potential use of a three-zone risk stratification scheme within a standardized image-based drawing framework. CONCLUSIONS: PD-MGMA-DSCNN provides state-of-the-art performance in handwriting-based PD detection while remaining a compact, systematically optimized model with clinically interpretable explanations. The framework may support tablet-based PD screening or triage in movement-disorder clinics and longitudinal monitoring, and serves as a practical starting point for incorporating digital handwriting biomarkers into clinical PD care.