Neural network estimation of balance control during locomotion

Abstract

Gait patterns of the elderly are often adjusted to accommodate for reduced function in the balance control system and a general reduction in skeletal muscle strength. Recent studies have demonstrated that measures related to motion of whole body center of mass (COM) can distinguish elderly individuals with balance impairment from healthy peers. Accurate COM estimation requires a multiple-segment anthropometric model, which may restrict its broad application in assessment of dynamic instability. Although temporal-distance measures and electromyography have been used in evaluation of overall gait function and determination of gait dysfunction, no studies have examined the use of gait measurements in predicting COM motion during gait. The purpose of this study was to demonstrate the effectiveness of an artificial neural network (ANN) model in mapping gait measurements onto COM motion in the frontal plane. Data from 40 subjects of varied age and balance impairment were entered into a 3-layer feed-forward model with back-propagated error correction. Bootstrap re-sampling was used to enhance the generalization accuracy of the model, using 20 re-sampling trials. The ANN model required minimal processing time (5 epochs, with 20 hidden units) and accurately mapped COM motion (R-values up to 0.89). As training proportion and number of hidden units increased, so did model accuracy. Overall, this model appears to be effective as a mapping tool for estimating balance control during locomotion. With easily obtained gait measures as input and a simple, computationally efficient architecture, the model may prove useful in clinical scenarios where electromyography equipment exists.

Introduction

As humans age, gait patterns are adjusted to accommodate for reduced function in the balance control system and a general reduction in skeletal muscle strength (Grimby and Saltin, 1983; Fiatarone and Evans, 1993). The temporal-distance (T-D) measures of gait have been widely used in evaluation of overall function and determination of gait dysfunction in the elderly (Heitmann et al., 1989; Ferrandez et al., 1990; Elble et al., 1991; Leiper and Craik, 1991; Judge et al., 1996; Maki, 1997; Menz et al., 2003). These studies showed that while T-D measures of gait do provide an overall impression of walking performance, there is substantial inter-subject variability in the measures. Such variability may contribute to a lack of power in accurately predicting the risk of falling in the elderly. The effect of aging on muscle activation and strength in the elderly has been shown to result in higher electromyographic (EMG) signal amplitudes during gait (Finley et al., 1969; Shiavi, 1985). However, the resulting force production in aged subjects is highly variable (Galganski et al., 1993; Grabiner and Enoka, 1995). None of the previous studies have examined the effect of T-D parameters and EMG activity on control of whole body stability.

Control of whole body stability has been studied in recent years by analyzing motion of the whole-body center of mass (COM), the patterns of which were reported to be quite consistent during locomotion (Jian et al., 1993; MacKinnon and Winter, 1993; Prince et al., 1994; Winter, 1995). More recent studies have demonstrated an ability to distinguish elderly individuals with balance impairment from their age-matched healthy peers, using measures of medio-lateral (M-L) COM motion during obstacle crossing (Chou et al., 2003; Hahn and Chou, 2003). Accurate estimation of the whole-body COM requires three-dimensional reconstruction of a multiple segment biomechanical model. This technical requirement alone may restrict broad application of assessing dynamic instability.

In many clinical settings, gait analysis can be performed with accuracy in measures of gait velocity, stride length, stride time and step width. Additionally, with many brands of inexpensive hardware/software currently available, the relative magnitude of muscle activations may be measured with surface EMG during locomotion and other activities of daily living. Using T-D and EMG data to predict dynamic stability would be advantageous by reducing the necessity for a multiple camera motion analysis system (more costly), and reducing the time commitment of data-processing and analysis. A model is therefore needed which would allow accurate description of whole-body balance control, given simple measures of gait such as EMG and T-D parameters.

One approach for mapping interactions between gait measurements and balance control is to construct a nonlinear model using an artificial neural network (ANN). Biological nervous systems are capable of learning by adjustment of the synaptic connections between individual neurons. The ANN is modeled in a similar fashion, allowing the network to be trained by exposure to a set of input data where the output values are known. Weights of the ANN interconnections are iteratively adjusted to attempt correction of the final processed output to match that of known values. Once a network has been trained to a satisfactory level, the knowledge gained by this learning process is stored in the connection weights (synapses), allowing a trained network to solve new problems similar to the task it was trained on. The primary advantages of ANNs in solving real-world classification problems are (1) their resilience in the face of noise and variability within a dataset, and (2) the ability to map relationships between variables that would not otherwise be noticeable. They have been used with high success in problems that are either too complex for conventional methods or are of an exploratory nature (Chau, 2001).

Applications of ANN models in musculoskeletal biomechanics have dealt primarily with joint angles and joint moment estimations in gait simulation (Sepulveda et al., 1993) and estimation of muscle recruitment in static conditions (Nussbaum et al., 1995). Sepulveda and colleagues used traditional back-propagation algorithms to successfully map the relationship between EMG and joint angles, and between EMG and joint moments during gait. Nussbaum et al. also reported success using a back-propagation algorithm to map lumbar muscle recruitment during moderate static exertions. Koike and Kawato (1995) used an architecturally complex ANN model to estimate isometric joint torques and trajectory from surface EMG in upper limb motions. More recent efforts by Luh et al. (1999) showed promising results with use of a simple, three-layer ANN, using an adaptive learning rate back-propagation algorithm in the determination of elbow joint torque from EMG activity.

Although ANN modeling has been used in studies of human locomotion (Holzreiter and Kohle, 1993; Sepulveda et al., 1993; Lafuente et al., 1998; Prentice et al., 1998; Savelberg and de Lange, 1999; Su and Wu, 2000; Prentice et al., 2001; Wu et al., 2001), no previously published work has addressed the ability of such models to map the interaction between basic gait measurements and descriptions of dynamic balance control. The purpose of this study was to demonstrate the effectiveness of an ANN model in mapping gait measurements (normalized lower extremity EMG signals and basic T-D parameters) onto whole body measures of dynamic stability (motion of the COM). It was hypothesized that a relatively simple ANN architecture would be capable of accurately mapping interactions between these variables.