Uncertainty-Enhanced Spatiotemporal Framework for Sit-to-Stand Temporal Segmentation in Parkinson's Disease

Automated temporal segmentation of sit-to-stand (STS) transitions in Parkinson's disease (PD) patients presents significant challenges due to multi-scale temporal dynamics, boundary ambiguity, and severe class imbalance in continuous video streams. Existing STS assessment approaches primarily focus on pre-segmented action recognition rather than frame-level temporal localization, limiting their applicability for long-term home-based monitoring. To address these limitations, we propose an uncertainty-enhanced spatiotemporal framework that integrates graph convolutional networks with uncertainty modeling for precise temporal action segmentation. Specifically, we develop a novel spatiotemporal architecture that synergistically combines graph-based spatial joint modeling with hierarchical multi-scale temporal feature pyramids to capture complex movement patterns across different time scales. Moreover, we introduce Gaussian boundary uncertainty modeling that transforms sharp binary transitions into smooth probabilistic representations, capturing the gradual and ambiguous nature of pathological movement boundaries. Furthermore, we design a dual-head architecture that jointly optimizes action segmentation and boundary uncertainty estimation to effectively handle severe class imbalance and inherent annotation ambiguity in clinical data. Additionally, we develop a boundary-aware post-processing mechanism that leverages learned uncertainty estimates for improved temporal localization. Comprehensive experiments on the real-world PD dataset demonstrate superior performance with 92.59% F1@50 (segment-level F1 at 50% overlap) and 81.78% mIoU (mean Intersection over Union), outperforming existing temporal action segmentation methods. Ablation studies validate the effectiveness of each component, with statistical analysis confirming consistent performance across 5-fold cross-validation. Our framework establishes a foundation for continuous, automated mobility monitoring in clinical telemedicine applications, enabling precise detection of STS transitions from standard video recordings without requiring manual segmentation.