An EMG-driven musculoskeletal model of the shoulder

This paper aims to develop an EMG-driven model of the shoulder that can consider possible muscle co-contractions. A musculoskeletal shoulder model (the original model) is modified such that measured EMGs can be used as model-inputs (the EMG-driven model). The model is validated by using the in-vivo measured glenohumeral-joint reaction forces (GH-JRFs). Three patients carrying instrumented hemi-arthroplasty were asked to perform arm abduction and forward-flexion up to maximum possible elevation, during which motion data, EMG, and in-vivo GH-JRF were measured. The measured EMGs were normalized and together with analyzed motions served as model inputs to estimate the GH-JRF. All possible combinations of input EMGs ranging from a single signal to all EMG signals together were tested. The 'best solution' was defined as the combination of EMGs which yielded the closest match between the model and the experiments. Two types of inconsistencies between the original model and the measurements were observed including a general GH-JRF underestimation and a GH-JRF drop above 90° elevation. Both inconsistencies appeared to be related to co-contraction since inclusion of EMGs could significantly (p<.05) improve the predicted GH-JRF (up to 45%). The developed model has shown the potential to successfully take the existent muscle co-contractions of patients into account.     

Determining a long term ambulatory load profile of the shoulder joint: neural networks predicting input for a musculoskeletal model

To gain more insight in the development of joint damage, a long term load profile of the shoulder joint under daily living conditions is desirable. Standard musculoskeletal models estimate joint load using kinematics and exerted force. However, the latter cannot be measured continuously in ambulatory settings, hampering the use of these models. This paper describes a method for obtaining such a load profile, by training a Neural Network (NN), using kinematics and EMG. A small data set of specified movements with known exerted forces is used in two ways. First, the model calculates several variables of joint load, and a set of Generalized Forces and Net Moments (GFNM) around the model's degrees of freedom. Second, using kinematics and EMG, an NN is trained to predict these GFNM, which can concurrently be used as input for the model, resulting in full model output independent of exerted force. The method is validated with an independent trial. The NN could predict GFNM within 10% relative RMS, compared to output of the model. The NN-model combination estimated joint reaction forces with relative RMS values of 7 to 17%, enabling the estimation of a detailed load profile of the shoulder under daily conditions.