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

Area of Science:

• Biomedical Engineering
• Computer Vision
• Neurology

Background:

• Automated sit-to-stand (STS) transition analysis in Parkinson's disease (PD) is crucial for home monitoring but faces challenges like complex temporal dynamics and ambiguous boundaries.
• Current methods often lack frame-level precision, hindering long-term patient observation and assessment in real-world settings.

Purpose of the Study:

• To develop an advanced spatiotemporal framework for accurate, automated temporal segmentation of STS transitions in PD patients.
• To address limitations in existing approaches by focusing on precise temporal localization and handling data challenges like class imbalance and boundary ambiguity.

Main Methods:

• Proposed an uncertainty-enhanced spatiotemporal framework integrating graph convolutional networks (GCNs) and uncertainty modeling.
• Developed a novel architecture combining GCNs for spatial modeling with multi-scale temporal feature pyramids.
• Introduced Gaussian boundary uncertainty modeling and a dual-head architecture for joint segmentation and uncertainty estimation, alongside boundary-aware postprocessing.

Main Results:

• Achieved superior performance on a real-world PD dataset with 92.59% F1@50 and 81.78% mIoU, outperforming existing methods.
• Ablation studies confirmed the effectiveness of individual framework components.
• Statistical analysis demonstrated consistent performance across 5-fold cross-validation.

Conclusions:

• The developed framework enables precise temporal action segmentation for STS transitions in PD patients, overcoming limitations of prior methods.
• Establishes a foundation for continuous, automated mobility monitoring in telemedicine, facilitating early detection and management of PD progression.
• Demonstrates the potential of uncertainty modeling in improving the accuracy of video-based clinical assessments.