Predicting Ankle Joint Moments Using A Hybrid EMG-driven Model

Abstract

To study sport medicine injuries and their effects on human motion, accurate biomechanical models are important. For these models to characterize the response to injury, they must take into account the differences in the ways injured people move their limbs and activate their muscles. We have created a biomechanical model of the ankle that uses subject specific muscle activation as an input and predicts the corresponding joint moment (plantar-flexion/dorsiflexion). PURPOSE To demonstrate the ability of our model to accurately predict joint moments given electromyograms (EMGs) and joint position data. METHODS Three types of data were collected during isokinetic and gait trials: (1) EMG from the tibialis anterior, medial gastrocnemius, lateral gastrocnemius, and soleus; (2) joint position, (3) and reaction forces (from the ground or dynamometer). Ankle joint moment was calculated in two ways: (1) forward dynamics using EMG and joint position data, and (2) inverse dynamics using joint position and reaction force data. The joint moment determined from the inverse dynamics calculation was used to calibrate the forward dynamic estimation of the moment. Adjustable parameters in the forward dynamics model were optimized to produce a best fit. Once calibrated, only the forward dynamics model was used to predict the joint moments from novel trials. RESULTS Preliminary data were collected from three test subjects. A comparison between the predicted and the measured ankle joint moment was performed; root mean squared error was approximately 10%, and R-squared values were of 0.95, 0.95, 0.87 for each subject respectively. CONCLUSION The curve shape was very closely matched in the prediction of concentric tasks, but the largest errors were observed during eccentric tasks. Further adjustment and refinement of the model parameters should correct this. One of the major strengths of the model is that is allows estimation of muscle forces (in contrast to inverse dynamics models) and relies on a subject's actual muscle activation values (in contrast to optimization approaches). Supported by NIH grant R01 HD38582.