ScholarWorks@성균관대학교

초록

Recent studies on automated sit-to-stand (STS) assessment for Parkinson’s disease (PD) have been evaluated almost exclusively in laboratory settings, which limits deployment in daily environments that involve varied movements and fine-grained severity differences. To address this gap, we propose Kinematic-Guided Assessment (KiGA), a novel dual-branch framework for video-based automated STS severity assessment in real-world conditions. The first branch is a dynamic graph-based spatiotemporal network that employs adaptive spatial learning and multi-scale temporal convolutions to capture the intricate, coordinated variations of joint motion. Running in parallel, the second branch aggregates symptom-driven kinematic features that are explicitly mapped to clinical criteria, thereby delivering robust, real-world estimates of four key motor domains—postural stability, body bradykinesia, leg rigidity, and arm rigidity. Additionally, we design a class-specific contrastive learning strategy to enhance spatiotemporal feature discrimination between severity levels despite environmental variability. Comprehensive experiments on real-world data demonstrate the effectiveness of our framework by achieving 77.94% accuracy and 89.68% acceptable accuracy while exhibiting consistent performance across different camera viewpoints, significantly outperforming existing approaches. Statistical analysis validates the clinical relevance of extracted kinematic features, with most showing high statistical significance (p<0.001). These results demonstrate the potential for reducing the burden on specialist neurologists and enabling continuous remote monitoring of PD patients in their home environments. Our framework provides an accessible tool for objective PD assessment using only standard video recordings, establishing a foundation for automated movement analysis in telemedicine settings.

키워드