Data Acquisition Through Participatory Design for Automated Rehabilitation Assessment

2025-01-02

AI Summary

Clinical assessment of upper-limb function in stroke patients relies heavily on subjective, manual administration of the Action Research Arm Test (ARAT), resulting in low efficiency and poor standardization. Method: This study proposes a semi-automated ARAT assessment system co-designed with clinicians and patients. It integrates multi-view unobtrusive video capture, an expert-informed yet non-expert-friendly interactive motion segmentation tool, and a structured clinician rating interface. We introduce a novel “clinical workflow-embedded” data collection paradigm to ensure training data align with clinical validity, operational feasibility, and expert intuition. Contribution/Results: Five physicians independently collected 1,800 high-quality videos (error rate <5%); three non-experts segmented 760 motion segments in ~20 seconds each, selecting optimal views for >90% of cases. Clinician feedback indicated minimal usability issues. The system significantly enhances objectivity, test–retest reliability, and clinical deployability of ARAT assessments.

Technology Category

• Computer Vision: Segmentation
• Humans and AI: Interaction Techniques and Devices
• Intelligent Robots: Manipulation

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

Through participatory design, we are developing a computational system for the semi-automated assessment of the Action Research Arm Test (ARAT) for stroke rehabilitation. During rehabilitation assessment, clinicians rate movement segments and components in the context of overall task performance. Clinicians change viewing angles to assess particular components. Through studies with clinicians, we develop a system that includes: a) unobtrusive multi-camera capture, b) a segmentation interface for non-expert segmentors, and c) a rating interface for expert clinicians. Five clinicians independently captured 1800 stroke survivor videos with<5$%$ errors. Three segmentors have segmented 760 of these videos, averaging 20 seconds per segment. They favor the recommended camera view $>$ 90%. Multiple clinicians have rated the segmented videos while reporting minimal problems. The complete data will be used for training an automated segmentation and rating system that empowers the clinicians as the ratings will be compatible with clinical practice and intuition.

Problem

• Stroke Rehabilitation
• Computer Technology
• Arm Function Assessment

Innovation

• Crowdsourced Data Collection
• Stroke Rehabilitation Assessment
• Automated Tool Training