| ObjectiveThe purpose of this study is to develop an artificial neural network model that can accurately predict vertical ground reaction force（vGRF）during running using a wearable inertial sensor（IMU）.vGRF is a key monitoring indicator in the study of running injuries,but it has long relied on force plate and is not easy to obtain.Recent advances in wearable sensor technology and artificial neural network models provide a promising approach for getting the vGRF data.To this end,this study aims to:（1）Propose a method for aligning and synchronizing motion data acquisition with IMU and instrumented treadmill.（2）Evaluate the performance of wavelet neural network model（WNN）and feed-forward neural network model（FFNN）based on single IMU triaxial acceleration data,and explore the potential of predicting vGRF based on IMU uniaxial（sagittal）acceleration data.The goal is to identify an artificial neural network model for predicting running vGRF based on acceleration data collected by IMU.MethodsFifteen runners participated in this study,during which shank triaxial acceleration data and vGRF values were collected at speeds of 12,14,and 16 km/h via a single IMU and instrumented treadmill.To align and synchronize the data,a novel time-window algorithm was proposed,which utilized the variability characteristics of the running gait cycle.The time of One or more running cycles was a time window.IMU and vGRF data curves were divided by a time window to obtain two sets of time window series;The intermediate time window was selected as the reference time window in the IMU time window series set,and the root mean square error（RMSE）was calculated with the vGRF time window series set to obtain a set of RMSE series;The IMU time window and vGRF time window corresponding to the minimum RMSE were the synchronization points,which were calculated and confirmed;Intercept all one-to-one synchronization data before and after the synchronization points;The average absolute and relative error,the maximum absolute and relative error,and the Bland-Altman method were used to calculate the error of the synchronization data to verify the accuracy and effectiveness of the synchronization algorithm.The foot strike angle was utilized to distinguish between forefoot strike（FFS）,midfoot strike（MFS）,and rearfoot strike（RFS）,with MFS and FFS being classified into a single class.One person’s data was randomly selected from the FFS and RFS runner data sets as the test set,while the remaining data was used as the training set.The Morlet and Sigmoid functions were used as the activation function of the three-layer WNN and three-layer FFNN,respectively.In this study,the model input was optimized,and the data of the entire stage of the running support period was used as a sample input to facilitate the model to identify complete information and find accurate rules.In addition,the loss function was optimized,and on the basis of the mean squared error loss function,L2norm regularization was added to alleviate the overfitting of the model.In this study,the models were evaluated using the coefficient of multiple correlation（CMC）,error values,and Bland-Altman method.ResultsUsing the time-window alignment and synchronization algorithm,the mean absolute error（ranging from 0.003 to 0.027 seconds）was less than 0.03 seconds,and the mean relative error（0.4%to 4.1%）was less than 5%for synchronized data from15 subjects.The Bland-Altman consistency results indicate that the mean errors of the step time were close to zero（ranging from-0.002 to 0.002 seconds）,and the maximum errors（ranging from 0.055 to 0.055 seconds）were within the 95%confidence interval.These results demonstrate that the time-window alignment and synchronization algorithm effectively achieves data synchronization between IMU and instrumented treadmill.Regarding the WNN model using shank triaxial or sagittal acceleration data via a single IMU,for both FFS and RFS runner data sets at speeds of 12,14,and 16km/h,the CMC between the predicted and measured vertical ground reaction force（vGRF）values exceeded 0.99,and the normalized root mean squared error（NRMSE）was less than 7%.The NRMSE and mean absolute percentage error（MAPE）between the predicted and measured peak vGRF values were less than 8%,and the Bland-Altman analysis demonstrated that the mean errors between the predicted and measured peak vGRF values were close to zero,with maximum errors falling within the 95%confidence interval.These results illustrate that the vGRF predicted curves and peak vGRF predicted values by the WNN models are highly similar and accurate compared to the measured curves and values.However,based on the FFNN model using shank triaxial or sagittal acceleration data via a single IMU for FFS and RFS runner data sets at speeds of 12,14,and 16 km/h,the CMC between the predicted and measured vGRF values was greater than 0.97,and the NRMSE ranged from 5.70%to 9.21%.Nevertheless,using the FFNN model with triaxial shank acceleration data for the RFS data set,the average errors between the predicted and measured impact and active peak vGRF values were slightly deviated from zero（ranging from-0.31 to 0.50 BW）,and the5.56%ratio of errors between the predicted and measured impact peak vGRF values exceeded the 95%confidence interval.Additionally,using the FFNN model with sagittal acceleration data for the FFS and RFS runner data sets,the ratios of errors between the predicted and measured active peak vGRF values at 16 km/h and impact peak vGRF values at 12 km/h（9.09%and 3.85%,respectively）exceeded the 95%confidence interval,while the ratio of errors between the predicted and measured active peak vGRF values at 14 km/h（11.11%）also exceeded the 95%confidence interval.These findings indicate that while the vGRF predicted curves by the FFNN models are highly similar to measured curves,the peak vGRF predicted values by the FFNN models exhibit moderate accuracy compared to measured values.ConclusionThe study proposes a new algorithm for running data synchronization between wearable sensors and measurement systems,which can effectively solve the synchronization problem.Additionally,the study develops the WNN model using shank sagittal acceleration data via a single IMU,which is a new method to accurately predict vGRF during running.This method has the characteristics of more accurate,stable,cheap and portable,and has important application effect on the prevention and rehabilitation of running injury. |
