An End-to-End AI-Assisted Workflow for Patient-Specific 3D Digital Human Reconstruction and Unreal Engine-Based Stroke Rehabilitation Simulation

This study develops and validates an end-to-end artificial intelligence (AI) workflow for patient-specific three-dimensional (3D) digital human modeling, anatomically constrained rigging, upper-limb motion reconstruction, and Unreal Engine-based stroke rehabilitation simulation. A prospective study involving 90 patients with stroke and 30 healthy controls was conducted. Multi-view images, anthropometric measurements, clinical assessments, and motion-capture data were integrated for model training and independent testing. The system's performance was evaluated based on reconstruction accuracy, processing efficiency, expert-rated clinical quality, and the associations between derived digital biomarkers and clinical upper-limb function. The proposed workflow achieved a structural similarity index measure (SSIM) of 0.958, a learned perceptual image patch similarity (LPIPS) of 0.083, an anatomical landmark error of 11.2 mm, and a Chamfer distance of 3.84 mm. Furthermore, the total processing time was reduced to 20.6 minutes, representing a 69.0% decrease compared to an automated baseline and a 91.0% decrease versus manual production. Clinical validation demonstrated that movement smoothness strongly correlated with the Action Research Arm Test (r = 0.73), while shoulder flexion range of motion correlated with the Fugl-Meyer Assessment for Upper Extremity (r = 0.72). In conclusion, this workflow provides accurate, efficient, and clinically interpretable digital-human reconstruction, establishing a robust foundation for objective stroke rehabilitation assessment.