Fall risk assessment via multi-batch plantar pressure data fusion: an ensemble framework integrating global and local features.

Introduction: Falls seriously threaten the health and independence of older adults. Existing plantar pressure-based fall risk assessment methods often rely on single datasets and global features, limiting generalizability across data distributions. Methods: To enhance cross-batch robustness, this study integrated plantar pressure data from two independent batches with inherent distributional differences. Eighty older adults performed a 2-min walking task while pressure signals were collected using an intelligent footwear system. A feature set of 44 global and 456 local features was constructed, and a four-model ensemble comprising two MLPs and two 1D-CNNs was developed. Model performance was evaluated using 8-fold stratified cross-validation, Mann-Whitney U-based pre-selection, and SHAP/Grad-CAM interpretability analyses. Results: The ensemble achieved an overall accuracy of 86.25%, outperforming the best single model (MLP1, 80.0%). Performance remained stable across the two batches, with accuracies of 86.0% and 86.7%, corresponding to a fluctuation of 0.7 percentage points. This fluctuation was substantially lower than those observed for the single models; for example, CNN2 fluctuated by 9.3 percentage points. Discussion: These findings suggest that fusing multi-batch plantar pressure data with an ensemble integrating global and local features can help mitigate the impact of cross-batch distribution shifts on model performance and improve the stability of fall risk assessment, thereby providing a methodological basis for its translation from research settings to practical applications.