Empirical evaluation of optimization-based lumbar muscle force prediction models.

Abstract

This dissertation investigates the ability of optimization-based models to predict torso muscle electromyographic (EMG) data during static, asymmetric tasks. Four optimization model formulations are evaluated: (1) a quadratic program which minimizes the sum of the cubed muscle contraction intensities; (2) a quadratic program which minimizes the sum of the squared muscle contraction intensities; (3) a double linear optimization model which minimizes spinal compression force using the lowest possible muscle contraction intensity bounds; and (4) a model based on central nervous system control of muscular "synergies." Experimental conditions which give distinct model predictions are used to distinguish between model formulations. Two experiments require subjects to statically resist combined sagittal and frontal plane moments about the L3/L4 vertebral level. The loading conditions are selected to provide model predictions which can be tested with limited assumptions about the force-EMG relationship. The data indicate that the double linear optimization scheme, the sum of squared muscle forces, and the synergy model predictions substantially deviate from the measured EMG data. The final experiment requires subjects to statically resist extension and twisting moments. The abdominal muscle EMG data do not support the double linear optimization model. The sum of cubed muscle intensities formulation appears to predict agonist muscle activity the best for all three experiments. Definitions of antagonism, which are based on a mechanical analysis of the biochemical system, are also developed. The definitions are stated in two ways: (1) as "primal" statements about the existence of muscle forces which oppose the net joint moment; and (2) as "dual" statements about the moment contributions of the muscles. Farkas's theorem is used to prove that the primal and dual forms are equivalent. It is proven that single joint models which have a strictly positive gradient for positive muscle forces do not predict co-contraction as defined here. Both the primal and dual definitions are used to classify the observed EMG activity. An experimental protocol for estimating muscle tension from surface EMG is also developed to facilitate the application of the primal form of the antagonism definition. The procedure applies principal components regression to EMG data recorded during isometric, anisotonic "ramp" contractions involving flexion, extension, and lateral bending. It is shown that the procedure provides better force-EMG parameter estimates than traditional multiple regression techniques. Muscle forces which are estimated from EMG data indicate that co-contraction adds 12% to 14% to L3/L4 spinal compression force over a wide range of asymmetric loading conditions.