STROKEVISION-BENCH: A MULTIMODAL VIDEO AND 2D POSE BENCHMARK FOR TRACKING STROKE RECOVERY

Area of Science:

• Neurorehabilitation
• Computer Vision
• Biomedical Engineering

Background:

• Clinical assessment of upper extremity (UE) function post-stroke is often subjective, limiting detection of subtle motor improvements crucial for rehabilitation.
• Existing datasets lack focus on structured clinical assessments like block transfer tasks, hindering objective analysis of stroke recovery.

Purpose of the Study:

• To introduce StrokeVision-Bench, the first dataset specifically designed for objective UE motor function assessment in stroke patients using block transfer tasks.
• To establish performance baselines for automated analysis of stroke rehabilitation using computer vision techniques.

Main Methods:

• Developed StrokeVision-Bench, a dataset of 1,000 annotated videos of stroke patients performing the Box and Block Test (BBT).
• Videos are categorized into four action classes and include raw frames and 2D skeletal keypoints.
• Benchmarked state-of-the-art video action recognition and skeleton-based classification methods.

Main Results:

• Successfully created a specialized dataset (StrokeVision-Bench) for stroke rehabilitation research.
• Established initial performance benchmarks for automated UE function assessment using established AI models.
• Demonstrated the feasibility of using computer vision for objective analysis of block transfer tasks.

Conclusions:

• StrokeVision-Bench addresses a critical gap in stroke rehabilitation datasets, enabling objective and scalable UE function assessment.
• Automated analysis of block transfer tasks holds significant potential for personalized stroke rehabilitation planning.
• This resource will facilitate future research in AI-driven neurorehabilitation.