Contraction-based variations in upper limb EMG-force models under isometric conditions

In this work, a previously developed model, which maps joint kinematic data and estimated muscle activation levels to net elbow joint torque, is trained with 4 groups of datasets in order to improve force estimation accuracy and gain insight into muscle behaviour. The training datasets are defined such that surface electromyogram (EMG) and force data are grouped within individual trials, across trials, within force levels and across force levels, and model performance is assessed. Average evaluation error ranged between 5% and 15%, with the lowest error observed for models trained with datasets grouped within separate force levels. Model error is further reduced when training datasets are grouped across data collection trials. Therefore, more accurate estimation of elbow joint behaviour can be accomplished by taking into account the functional requirements of muscle, and allowing for separate models to be developed accordingly.