Can shoulder joint reaction forces be estimated by neural networks?

To facilitate the development of future shoulder endoprostheses, a long term load profile of the shoulder joint is desired. A musculoskeletal model using 3D kinematics and external forces as input can estimate the mechanical load on the glenohumeral joint, in terms of joint reaction forces. For long term ambulatory measurements, these 3D kinematics can be measured by means of Inertial Magnetic Measurement Systems. Recording of external forces under daily conditions is not feasible; estimations of joint loading should preferably be independent of this input. EMG signals reflect the musculoskeletal response and can easily be measured under daily conditions. This study presents the use of a neural network for the prediction of glenohumeral joint reaction forces based upon arm kinematics and shoulder muscle EMG. Several setups were examined for NN training, with varying combinations of type of input, type of motion, and handled weights. When joint reaction forces are predicted by a trained NN, for motion data independent of the training data, results show a high intraclass correlation (ICC up to 0.98) and relative SEM as low as 3%, compared to similar output of a musculoskeletal model. A convenient setup in which kinematics and only one channel of EMG were used as input for the NN's showed comparable predictive power as more complex setups. These results are promising and enable long term estimation of shoulder joint reaction forces outside the motion lab, independent of external forces.