DS-HViT: A Dual-Stream Hierarchical Vision Transformer for Multi-Scale Analysis of Parkinson's Disease Handwriting

Résumé

Early detection of Parkinson's Disease (PD) remains challenging in the medical field due to motor symptoms manifesting across different temporal scales. To address this, we introduce a Dual-Stream Hierarchical Vision Transformer (DS-HViT), a deep learning framework designed to capture the multi-scale temporal dynamics of PD-related Handwriting (HW) impairments. The architecture employs parallel micro- and macro-scale streams: the micro-scale stream models high-frequency tremor signatures, while the macro-scale stream captures gradual motor decline. The model was evaluated on the NewHandPD dataset using 5-fold patient-stratified cross-validation, complemented by bootstrap analysis and McNemar’s test for statistical validation. DS-HViT achieved an accuracy of 98.2 ± 0.8%, with sensitivity of 97.3 ± 1.2% and specificity of 99.0 ± 0.9%, significantly surpassing state-of-the-art methods (p < 0.01). Ablation studies confirmed the synergistic effect of dual-stream processing, while the model demonstrated excellent calibration with an Expected Calibration Error (ECE) of 0.043 and exceptional discriminative ability with a Diagnostic Odds Ratio (DOR) of 486.3.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0