Subject-Independent Estimation of Vertical Ground Reaction Force from a Wearable Accelerometer: A Reproducible Benchmark of Accuracy, Robustness, and Inference Latency

Accurate estimation of vertical ground reaction force (vGRF) outside laboratory environments remains a key challenge in wearable biomechanics and rehabilitation monitor- ing. This study investigates subject-independent vGRF estimation from a single triaxial tibial accelerometer using a fully repro- ducible experimental pipeline. A leave-one-subject-out (LOSO) validation protocol was employed to evaluate cross-subject gener- alization. Classical models (Ridge, ARIMAX, ensemble methods) and temporal neural networks (LSTM, GRU, TCN, CNN1D) were compared under identical preprocessing and windowing conditions. Robustness was assessed using controlled signal perturbations, including additive noise, axis ablation, spatial rotation, and temporal occlusion. Statistical differences among models were evaluated using the Friedman test and Kendall's concordance coefficient (W). Results indicate that nonlinear models-particularly Random Forest and GRU-achieve the highest predictive accuracy under subject-independent condi- tions, whereas temporal convolutional architectures provide com- petitive performance with lower inference latency. Robustness analyses reveal sensitivity to temporal occlusion and vertical-axis information, highlighting biomechanical dependencies in force reconstruction. All code and experiments are publicly available to ensure full reproducibility. These results provide practical guidelines for selecting deployable wearable models for gait monitoring and rehabilitation applications.