Environment-sensitive motion modelling in healthcare with synthetic retargeting

To address the critical data scarcity and privacy constraints that limit video-based motor behaviour assessment in clinical settings through a synthetic data generation framework, enabling robust human detection with high fidelity across challenging scenarios.

We employed synthetic data generation tailored to specific environments, implementing a novel synthetic retargeting approach based on procedural image synthesis. This method addresses the critical obstacles of limited training data in clinical settings due to privacy concerns, constrained views, occlusions, and uncontrolled environmental characteristics.

Our synthetic retargeting approach yielded substantial and statistically significant performance improvements in human detection under real-world clinical data regimes. Evaluated across two clinical scenarios, the method improved existing models' performance (human detection score) by up to 19.4% in the more challenging scenario and up to 9.8% in the less challenging scenario (both with p < 0.001), demonstrating both high fidelity and robustness against challenging environments.

Synthetic retargeting provides an efficient and effective solution for adapting pre-trained human detection models to specific clinical deployment scenarios by generating scenario-tailored synthetic data, circumventing the privacy and logistical constraints that limit real data collection in healthcare settings. This approach enables robust video-based motor behaviour quantification with significant implications for both clinical management and research.