Hand Grasp Classification in Egocentric Video After Cervical Spinal Cord Injury.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Human-Computer Interaction

Background:

• Hand function is critical for independence in individuals with spinal cord injury (SCI).
• Wearable cameras offer a non-clinical method for assessing hand function.
• Analyzing hand grasps and frequency is vital for precise clinical evaluation.

Purpose of the Study:

• To develop a hierarchical model for summarizing hand grasping strategies in SCI individuals at home.
• To enable precise analysis of dominant hand grasps and usage frequency from video data.

Main Methods:

• A two-level hierarchical model was introduced for classifying hand-object interactions.
• The first level used hand-object contact estimation.
• The second level employed a deep model incorporating hand postures and contextual information.

Main Results:

• The first hierarchical level achieved 86% ±1.0% accuracy.
• Grasp classification accuracy averaged 66.2% ±12.9%, with significant participant variability (41%-78%).
• Optimal performance was achieved using a ResNet50 backbone with smoothed grasp classification.

Conclusions:

• A novel algorithm enables home-based analysis of hand movements in SCI patients.
• This technology can assist clinicians in evaluating hand function and potentially aid in other neurological conditions affecting hand use.