Deep Learning-Based Subtask Segmentation of Timed Up-and-Go Test Using RGB-D Cameras.

Area of Science:

• Biomechanics
• Medical Technology
• Artificial Intelligence

Background:

• The Timed Up-and-Go (TUG) test is a standard measure of functional mobility.
• Analyzing individual TUG subtask timings offers deeper clinical insights than total completion time.
• Existing TUG segmentation methods lack accuracy, efficiency, and objectivity.

Purpose of the Study:

• To develop a novel deep learning-based method for accurate and objective subtask segmentation of the TUG test.
• To evaluate the proposed method's performance across different subject groups (healthy young, healthy adults, stroke patients).
• To identify optimal input data for real-time TUG subtask analysis.

Main Methods:

• A dilated temporal convolutional network (TCN) was employed for TUG subtask segmentation.
• The system utilized a single RGB-D camera for data acquisition.
• Performance was evaluated using three distinct subject groups, comparing the proposed method against existing techniques.

Main Results:

• The deep learning approach achieved high segmentation accuracy across all groups: healthy young (95.46%), healthy adults (94.53%), and stroke patients (93.58%).
• The proposed method demonstrated superior generality and robustness compared to rule-based and traditional ANN methods.
• Using pelvis-derived input alone yielded the best accuracy, enabling real-time inference at approximately 15 Hz on edge devices.

Conclusions:

• The proposed deep learning-based TUG subtask segmentation method provides accurate, efficient, and objective functional mobility assessment.
• Pelvis-based input is sufficient for high-accuracy, real-time TUG analysis on edge devices.
• This technology has the potential to significantly aid clinical interventions and patient recovery monitoring.