Zhou Yulin, Fu Shengxing, Yao Tianqi, Liu Hui, Li Hanjun
School of Sport Science, Beijing Sport University, Beijing, China.
(b)Sports and Health Science Research Center, National Institute of Sports Medicine, Beijing, China.
J Biomech. 2025 Mar;181:112548. doi: 10.1016/j.jbiomech.2025.112548. Epub 2025 Jan 27.
Artificial neural networks (ANNs) offers potential for obtaining kinetics in non-laboratory. This study compared the estimation performance for ground reaction forces (GRF) and lower-limb joint moments during sidestepping between ANNs fed with full-body and lower-body landmarks. 71 male college soccer athletes executed sidestepping while three-dimensional kinematics and kinetics were collected to calculate joint moments by inverse dynamic. To estimate GRF and lower-limb joint moments, coordinates of 18 full-body (the full-body landmarks ANN) and 11 lower-limb body landmarks (the lower-body landmarks ANN) were respectively used as inputs in ANNs. Estimation performance was evaluated using the coefficient of multiple correlations, root mean square error (RMSE), and normalized RMSE (nRMSE) between estimated and measured results. A Wilcoxon signed-rank test determined the difference in estimation performance between the two types of ANNs. Statistical parametric mapping determined the difference between the estimated and measured curves. The lower-body landmarks ANN showed lower error for sagittal knee moments (RMSE: p < 0.001; nRMSE: p < 0.001), but higher error for sagittal hip (RMSE: p = 0.015) and ankle moments (RMSE: p = 0.001; nRMSE: p = 0.001). Significant differences between the lower-body landmarks ANN estimates and measurement curves were found in anterior-posterior GRF (10-12 %, p = 0.013), vertical GRF (5-15 %, p < 0.001), and hip transverse moment (1 %, p = 0.017). No significant differences were found in the estimated and measured GRF peaks. The ANN only using lower-body landmarks as inputs could accurately estimate GRF and lower-limb joint moments during sidestepping, with better performance for knee moments, while ANN using full-body landmarks performs better for hip and ankle moments.
人工神经网络（ANNs）为在非实验室环境中获取动力学数据提供了潜力。本研究比较了使用全身和下肢标志点输入的人工神经网络在侧步过程中对地面反作用力（GRF）和下肢关节力矩的估计性能。71名男性大学足球运动员进行侧步动作，同时收集三维运动学和动力学数据，通过逆动力学计算关节力矩。为了估计GRF和下肢关节力矩，分别将18个全身标志点（全身标志点人工神经网络）和11个下肢标志点（下肢标志点人工神经网络）的坐标用作人工神经网络的输入。使用估计结果与测量结果之间的多重相关系数、均方根误差（RMSE）和归一化均方根误差（nRMSE）来评估估计性能。采用Wilcoxon符号秩检验确定两种类型人工神经网络在估计性能上的差异。统计参数映射确定估计曲线与测量曲线之间的差异。下肢标志点人工神经网络在矢状面膝关节力矩方面显示出较低的误差（RMSE：p < 0.001；nRMSE：p < 0.001），但在矢状面髋关节（RMSE：p = 0.015）和踝关节力矩方面误差较高（RMSE：p = 0.001；nRMSE：p = 0.001）。在前后GRF（10 - 12%，p = 0.013）、垂直GRF（5 - 15%，p < 0.001）和髋关节横向力矩（1%，p = 0.017）方面，发现下肢标志点人工神经网络估计值与测量曲线之间存在显著差异。在估计的和测量的GRF峰值方面未发现显著差异。仅使用下肢标志点作为输入的人工神经网络能够准确估计侧步过程中的GRF和下肢关节力矩，在膝关节力矩方面表现更好，而使用全身标志点的人工神经网络在髋关节和踝关节力矩方面表现更好。
