A dual-branch vision framework for interpretable balance assessment via pose dynamics and Vision-LLM semantics

Reliable balance assessment is crucial for detecting fall risk, yet traditional methods are subjective and time-consuming, while current vision-based approaches often lack clinical interpretability. To address this, we present a dual-branch 3D vision framework that combines quantitative pose dynamics with clinician-inspired semantics using a single depth camera. During the Timed-Up-and-Go (TUG) task, our system captures skeletal data and applies a Multi-View Weighted Fusion Selection to identify 20 key spatiotemporal features from 413 candidates. Simultaneously, a pretrained Vision-Large Language Model (Vision-LLM) extracts ten clinically relevant semantic descriptors through prompt-based inference, eliminating the need for domain-specific fine-tuning. These features are concatenated into a 30-dimensional embedding to drive fall-risk classification and Berg Balance Scale (BBS) score regression via a Random Forest model. Experimental results demonstrate that our method achieves a 94.2% classification accuracy and an R 2 of 0.9604 for BBS prediction, along with a 94.4% trajectory agreement with gold-standard motion capture systems. Furthermore, the framework reduces assessment time by 80% compared to traditional clinical protocols. By bridging fine-grained kinematics with observable behavioral semantics, this automated and contactless approach demonstrates potential for objective, interpretable, and efficient balance evaluation.