Patient-Level Biomechanical Pose Encoding and Temporal Gait Modeling for Pediatric EVGS Assessment

<jats:p>Pediatric gait assessment is clinically important for children with cerebral palsy, but manual visual scoring requires expertise and can be difficult to scale. This report describes our solution for the CVPR 2026 Children Gait Challenge, where the goal is to predict bilateral Edinburgh Visual Gait Score (EVGS) labels and cerebral palsy gait subtypes from multi-view 2D pose sequences. We use a two-stage design. First, a compact temporal convolutional network models fixed-length multi-view pose tensors for left and right gait subtype prediction. Second, a patient-level clinical feature ensemble predicts the 34 bilateral EVGS labels. The feature representation summarizes joint angles, trunk and pelvis posture, foot geometry, cadence proxies, frequency-domain ankle motion, left-right symmetry, and stance/swing phase statistics. A Gradient Boosting, Extra Trees, and Random Forest ensemble is trained with repeated cross-validation and per-label threshold tuning. The final selected submission achieved a public leaderboard score of 0.54177. The released code, model artifacts, feature tables, and selected submission are publicly available for reproduction.</jats:p>