A back-propagation neural network model of lumbar muscle recruitment during moderate static exertions.

A model employing artificial neural networks (ANNs) is developed for the prediction of lumbar muscle activity in response to steady-state static external moment loads. The model is constructed using standard feedforward networks and trained with available data using the standard back-propagation algorithm. Training with a limited set of exemplars allowed accurate prediction of muscle activity for novel moment loads (generalization). Sensitivity analyses during training and testing phases showed that the choice of specific network parameters was not critical except at extreme values of those parameters. Model predictions were better correlated with experimental data than predictions made using two optimization-based methods (average r2 = 0.83 using ANNs and 0.65 using optimization). The results suggest that lumbar muscle response varies smoothly and consistently with respect to the magnitude and orientation of external moments, and they also imply an upper limit on the accuracy of muscle activity prediction using only moment loads as input. ANNs present a useful alternative to EMG- and optimization-based approaches by being both 'reality-based' and predictive.