MSTG-Transformer: Multivariate Spatial-Temporal Gated Transformer Model for 3D Skeleton Data-based Fall Risk...

Area of Science:

• Gerontology
• Biomedical Engineering
• Computer Science

Background:

• Falls in older adults represent a major global public health issue.
• Effective fall risk prediction is crucial for preventative interventions.
• Current methods struggle with complex skeletal data during ambulation.

Purpose of the Study:

• To develop a novel, data-driven approach for predicting fall risk in older adults.
• To leverage 3D skeleton data and advanced deep learning for enhanced prediction accuracy.
• To address challenges in multidimensional feature extraction from skeletal keypoints.

Main Methods:

• Developed the Multivariate Spatial-Temporal Gated Transformer (MSTG-Transformer) using preprocessed 3D skeleton data.
• Constructed gait cycle sequences to represent movement patterns and enhance group distinctions.
• Employed convolutional modules for spatial/topological feature extraction and a dual-stream encoder for spatio-temporal encoding.
• Utilized a voting scheme to integrate segment classifications for final fall risk determination.

Main Results:

• The MSTG-Transformer achieved a superior prediction accuracy of 0.9510 ± 0.0240 on a real-world dataset.
• The proposed approach outperformed classical fall risk prediction methods.
• Demonstrated the significance of skeletal keypoint interactions in accurate fall risk assessment.

Conclusions:

• The MSTG-Transformer offers a highly accurate and effective method for fall risk prediction in older adults.
• Integrating spatial, temporal, and topological features from 3D skeleton data is vital for robust prediction.
• This approach has the potential to significantly improve fall prevention strategies for the aging population.