Estimating ground reaction forces and moments during gait from multiple kinematic variables using a feedforward neural network

Background Ground reaction force (GRF) and ground reaction moment (GRM) are critical in gait analysis. While force plates provide accurate measurements, they are costly and spatially limiting. Objective This study aimed to evaluate the reliability and accuracy of GRF and GRM predictions using a feedforward neural network (FNN) integrating infrared camera-based positional data with accelerometer (ACC) data from wearable devices. Methods Eighty participants walked at their usual pace along a 10-meter walkway over force plates. Positional and ACC data of body segments were used to train the FNN to predict GRF and GRM. Prediction accuracy was assessed using multiple metrics, including root mean square error (RMSE) and normalized RMSE (NRMSE). Results Combining positional and ACC data improved GRF prediction in all directions (vertical, anterior-posterior, medial-lateral). The combined dataset achieved a correlation coefficient of 0.979 for medial-lateral GRF and an NRMSE of 6.07%. GRM predictions also benefited from ACC integration, especially in the sagittal plane, where R 2 reached 0.939, outperforming other models. The vertical direction and transverse axis yielded the lowest RMSE and NRMSE. Conclusions These findings surpass many previously reported results, demonstrating the superior performance of the proposed model compared with current state-of-the-art methods. The approach offers a cost-effective, flexible alternative to traditional force plates for clinical and sports assessments.