Estimation Of Joint Torque During Countermovement Jump From Position Coordinates Using Deep Residual Recurrent Network: 1058 Board #184 May 27 1:30 PM - 3:00 PM

Mechanical outputs (MO) exhibited explosively in the lower limbs are important in many sports. Vertical countermovement jump (VCMJ) is often utilized to evaluate the ability to exhibit MO. To calculate MO, inverse dynamics is performed with body position data and ground reaction forces (GRFs) recorded by the motion-capture system and force-plates. However, it is difficult to obtain GRFs without laboratory setting because force-plates are usually quite expensive. Because of this device-dependent issue, it is hard to obtain MO in the common sporting scenes. PURPOSE: To create and develop an artificial neural network that is possible to estimate MO without force-plates, with body position data as inputs. METHODS: We designed a deep residual recurrent network (DRRN) to estimate the sagittal right knee torque as MO. Datasets were established for training and evaluating DRRN. Eighteen young males performed VCMJ under 3 conditions (make counter movement freely, deeply and shallowly) with arm swing. Body position data and GRFs were recorded by motion-capture system (250Hz) and force-plates (1250Hz). Three out of 18 subjects’ data were randomly chosen as validation data (validation subject A, B, and C). The other 15 subjects’ data were divided into two groups, i.e., 80% for training data and 20% for test data. As the objective variable, sagittal right knee torque was calculated using inverse dynamics. Explanatory variables were sagittal body position data. Parameters of DRRN were determined by an optimization calculation that aimed to reduce the difference between actual and estimated torque. To evaluate the predictive performance of DRRN, R² score (R²) and root mean square error (RMSE) were calculated. RESULTS: R² and RMSE of whole validation data were 87.7%±6.2 and 0.23±0.07, respectively. These indicators suggest DRRN model has a consistency in the level of predictive ability. R² and RMSE of validation subject A, B and C were 82.8%±7.1 and 0.26±0.07, 89.2%±3.6 and 0.25±0.06, 91.2%±3.5 and 0.18±0.06, respectively. These differences among subjects suggest that personal characteristics might not have been processed sufficiently. CONCLUSIONS: Deep RRN is effective in the estimation of joint torque with only body position data as inputs.