Descriptive vs. machine-learning models of vastus lateralis in FES-induced knee extension

The aim of this study was to compare the predictive performance of pure machine-learning models of muscles under functional neuromuscular electrical stimulation (FES) to that of descriptive Hill type models incorporating various levels of machine-learning in some of their elements. Inputs to the models were FES pulse width and vastus lateralis length and velocity, while the output was the vastus lateralis contractile force. Three types of models were developed for comparison purposes: 1) a Hill-based descriptive model without machine-learning elements; 2) 2 types of Hill-based models with several machine-learning elements; and 3) pure machine learning models using multilayer perceptron (MLPs) and adaptive neurofuzzy inference systems (ANFIS), The results revealed that the pure descriptive Hill model and two of the pure machine learning model configurations were the most inadequate in modeling electrically stimulated muscle. On the other hand, mixed models (i.e., Hill models that incorporated several machine learning elements), yielded the best results, giving mean force prediction errors of less than 3.3 % for the testing set