Performance evaluation of an ensemble neural network system of estimating transtibial prosthetic socket pressures during standing, walking and condition perturbation.

Abstract

Providing suitable prosthetic sockets for the restoration of function following lower-limb amputation remains a significant issue in medical device prescription. Poorly designed sockets are associated with discomfort, poor quality function and injury, with quality linked to the capability of the socket to adequately distribute the forces from ambulation. Despite this link, systems of measuring stump-socket interface pressure have not seen use in clinical practice, in part due to limitations in functional performance. A technique using neural networks to relate external socket deformation to the internal pressure distribution was recently developed: this method has several advantages over contemporary systems but had not been evaluated in detail in dynamic situations. A wireless system estimating transtibial socket pressure distribution was produced. When supplied with simulated socket loads, an estimate produced from a group of networks (an ensemble) demonstrated improved accuracy and reduced variance. Work was undertaken to identify optimal design in terms of input data conditioning and post-estimate correction. This demonstrated that these can provide significant accuracy and reliability improvements. Measurements were taken from two transtibial amputees during standing, walking, walking on slopes, walking with coronal plane misalignment and walking with an alternative socket liner. An evaluation of the contributions to variance confirmed the applicability of ensembles in this application. The system proved capable recording significant differences in socket load distribution between different prosthesis configurations. For future investigation, this demonstrates that the technique is sensitive enough to examine the changes in the application of force which are present during daily use, device set-up and common fault conditions. The results of this study support further development of the practical aspects of the system, future work in producing a realistic load training system and extrapolation of results to other sockets, structures and engineering problems.