Artificial Neural Network modelling for studying posture

University Of Vermont &St Agric College, Burlington VT

Investigators

Abstract

A major challenge with the application of artificial neural network (ANN) modeling in examining the motor-sensory relation in postural control is that the neural synaptic weights that relate the inputs to the outputs of the ANN are chaotic in nature. In an earlier study through theoretical analysis, numerical simulations and experimental tests, the PI and her colleague have found that these weights are interdependent. The product of these weights is a statistically stable variable and can be used to quantify the input-output relation of the ANN, called Q value. The objective of this proposed research is to extend the above work to the area of human postural control. Specifically, we will explore whether or not a Q value concept in an ANN can be used to quantify the motor-sensory relation in a classical postural control task - maintaining upright balance when the supporting base is suddenly rotated in a toes up direction. We will construct an ANN model that includes two outputs and seven inputs. The two outputs are the EMG signals from ankle dorsiflexor and plantartlexor in response to the onset of the supporting base rotation. The seven inputs are average eye-target distance (distance from eye center to a visual target), head acceleration (both linear and angular), ankle joint rotation, ankle joint rotation speed, and ground reaction forces (both normal and shear) under feet. These inputs represent the mechanical stimulation to the visual, vestibular, and somatosensory systems, respectively. These inputs and outputs variables will be measured directly from two groups of elderly subjects: peripherally neuropathic and normal, non-peripherally neuropathic. We will then determine the weights in the ANN model by a backward-propagation training routine, and the corresponding Q values relating each output to each of the inputs. We will statistically compare the Q values among the multiple sensory inputs within each group. We hypothesize that under this experimental condition, the Q values relating postural muscle activities to the somatosensory inputs would be: (1) significantly higher than the Q values relating to other sensory inputs (such as visual and vestibular inputs) in normal, non-neuropathic subjects; and (2) significantly lower than the Q values relating to other sensory inputs in neuropathic subjects. It is hoped that this study will contribute to our understanding of how sensory information is used to control postural muscle activities, and how a modification in the motor-sensory relation can result in increased postural stability or falls.