Gait stability prediction through synthetic time-series and vision-based data

Gait stability assessment in older adults is challenged by limited data availability and measurement complexity, particularly among vulnerable populations and in limited resource settings. We address three research questions: (1) can synthetic data accurately replicate the statistical properties of gait parameters in older adults? (2) how effectively do synthetic data-trained models predict the Margin of Stability (MoS) when tested on real-world data? and (3) what specific biomechanical features contribute most significantly to the MoS predictions in older adults? To address these challenges, the present study proposes a novel approach to gait stability prediction by integrating computer vision with a data-centric synthetic data generation (SDG) approach using accessible, low-cost technology.

Using a public dataset from 14 healthy older adults (86.7 ± 6.2 years), we implemented a constraint-based SDG methodology that preserved biomechanical relationships through SDG metadata configuration and rank correlation-based constraints. Gait analysis was performed through a smartphone (Motorola Moto G5 Play) and the open-source MediaPipe algorithm to extract body landmarks from frontal plane gait videos, making the approach suitable for resource-limited settings.

Our approach achieved exceptional fidelity (97.09% overall) and maintained biomechanical variable relationships. The model trained exclusively on synthetic data (TSTR) outperformed the model trained on real data (TRTR), with error reductions (RMSE decreased by 56.3%, MAE by 58.2%, and MSE by 80.9%) and improved variance explanation (R ² increase of 31.2%). SHAP analysis revealed that the synthetic data approach enhanced feature attribution alignment with established principles, particularly for step width, BMI, and fall history.

Therefore, our results show that: (1) synthetic data accurately replicated gait parameters with high fidelity; (2) synthetic data-trained models outperformed real data-trained models in MoS prediction; and (3) step width, BMI, and fall history were the most significant predictors of MoS in older adults. These findings demonstrate the potential of synthetic biomechanical time series to overcome data scarcity, improve predictive modeling capabilities, and enhance clinical gait assessment through accessible, low-cost computer vision methods.