Single-Camera AI-Based Classification of Clinically Defined Grasp Patterns in Children with Cerebral Palsy

Camera-based sensing may provide a scalable approach to quantifying hand function in children with cerebral palsy (CP), but fine grasp patterns remain challenging because of occlusion, small anatomical structures, and complex hand–obobject interactions. This study investigated a single-camera AI-based framework for classifying the developmentally ordered grasp patterns of the Quality of Upper Extremity Skills Test (QUEST). Grasp-of-Cereal and Grasp-of-Cube tasks performed by 41 children with CP were recorded with a single RGB camera and analyzed using MediaPipe hand landmarks, SAM2-based hand segmentation, YOLOv9 object localization, and a 3D DenseNet-121 classifier. Trial-level accuracy, severity-weighted accuracy (SWA), and quadratic weighted kappa (QWK) were estimated with participant-level bootstrap 95% confidence intervals (CIs). Accuracy was 0.84 (95% CI, 0.750–0.914) for Cereal and 0.85 (0.795–0.901) for Cube; SWA was 0.922 and 0.907, and QWK was 0.579 and 0.756, respectively. Therapist inter-rater QWK (0.883 and 0.838) is reported as a descriptive reference only. Cereal macro-F1 (0.534) and balanced accuracy (0.494) indicated uneven class performance. Development and test recordings were temporally separated but not participant-independent; for Cube, 39 of the 41 children contributing development data also contributed follow-up test recordings. These findings support the feasibility of longitudinal within-cohort grasp classification but do not establish generalization to previously unseen children.