Estimation of Ground Reaction Force with Wearable Accelerometers and Deep Learning: Comparative Analysis between Temporal Architectures

Introduction This work proposes and evaluates models based on deep neural networks for predicting ground reaction force (GRF) from accelerometer signals coming from wearable inertial sensors. Methodology Three architectures were developed and compared: Bi-LSTM, TCN, and a hybrid TCN-BILSTM architecture. The methodology included cross-correlation analysis for selecting the most relevant sensor, signal pre-processing, temporal segmentation, and model training. Results The results demonstrated that the Bi-LSTM and Hybrid architectures presented the best performance, reaching high coefficients of determination (R2) and low errors (RMSE and rRMSE), while the TCN model presented shorter training time but with inferior predictive performance. Conclusion This study highlights the feasibility of using low-cost inertial sensors to estimate GRF outside the laboratory environment, contributing to applications in rehabilitation, sports monitoring, and digital health.