PDFormer: A Multimodal Transformer for Quantifying Motor Symptom Severity in Parkinson's Disease

Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by deteriorating motor function and remains incurable. Diagnosis and monitoring remain largely clinical, as the pathological gold standard - post-mortem examination - is inherently impractical in routine care; the scarcity of movement disorder specialists further underscores the need for scalable, objective assessment. In this study, we present a data-driven AI framework for automated and objective quantification of PD motor symptom severity. Specifically, we introduce the MVPD dataset, a clinically realistic video corpus encompassing 18 motor symptoms evaluated under the Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS) Part-III, filling a critical gap in publicly available data for PD motor symptoms severity evaluation. We propose PDFormer, a simple yet efficient multimodal Transformer that integrates DINOv2 appearance features with skeleton trajectories via a multi-scale hybrid self-attention module, enabling the capture of subtle motor cues and sustained bradykinesia across spatial-temporal scales. Across public and in-house datasets, PDFormer achieves state-of-the-art performance in calculating MDS-UPDRS scores, setting a new benchmark performance in this field. These results highlight the potential of PDFormer to standardize PD motor symptoms severity assessments and support precision PD progression management. Code will be released at: this https URL.