Artificial neural network models for analysis of lumbar muscle recruitment during moderate static exertions.

Abstract

Understanding the etiology of low back pain and injury is hampered by the difficulty in ascertaining force levels in the lumbar muscles. In the contemporary literature, two primary types of muscle-activity prediction models have been developed: electromyographic-based and optimization-based. Each of these methods has associated advantages, yet their limitations provide the motivation for the current work. This dissertation provides a new approach for the development of analysis tools and predictive torso muscle models using artificial neural networks. Three different types of models were developed, evaluated, and validated for tasks involving asymmetric static moment loads of 10-50 Nm. In the first modeling scheme, experimental data were used to create models that estimate a set of muscle activity levels in response to external load magnitudes. The results demonstrated that simple models, composed of few processing units, could estimate muscular activity over a wide range of exertion levels, and that muscular activity appears primarily driven by and therefore predictable from the magnitudes of applied loads in static situations. This model overcomes two limitations of contemporary predictive models in that it is easily validated using myoelectric measures and predicts antagonistic co-contractile activity more realistically than optimization-based methods. The model failed, however, to accurately estimate the activity of the Latissimus Dorsi. A second type of model employed competitive processes to discriminate the extent to which individuals differ in their muscle response patterns. Individuals appeared 'clustered' around several different response patterns, suggesting that subsets of subjects may have different strategies for muscle recruitment. Competition was also intrinsic to the third type of model, used to simulate muscle response in the absence of exemplars or prototypes. The predicted response patterns were well correlated with experimental data. The success of the simulation indicates that a consistent recruitment plan that incorporates competitive processes between muscles may exist and that minor variations of this plan can mirror inter-individual differences. This work emphasizes that muscle recruitment can be achieved through the use of local controls and processes. Despite success in muscle activity prediction, work must continue to develop algorithms with more physiologic veracity.