CLASSIFICATION OF SOFTWARE CONTROL ARCHITECTURES FOR A POWERED PROSTHESIS THROUGH CONVENTIONAL GAIT ANALYSIS USING MACHINE LEARNING APPLICATIONS

Abstract

The powered prosthesis for people with transtibial amputation offers the opportunity to more appropriately restore gait functionality with benefits, such as powered plantar flexion. In particular, various software control architectures provide unique capabilities for regulating the powered prosthesis during gait. One highly novel approach applies the winding filament hypothesis, which enables an advanced modeling of muscle characteristics, such as through introducing the attributes of titin into the muscle model. The objective of the research is to contrast the conventional control architecture of the BiOM-powered prosthesis compared with the winding filament hypothesis control architecture through machine learning classification. Four machine learning algorithms are applied through the Waikato Environment for Knowledge Analysis (WEKA): J48 decision tree, [Formula: see text]-nearest neighbors, logistic regression, and the support vector machine. The feature set is derived from the force signal acquired from a force plate, which is a conventional gait analysis system. The feature set applied five attributes representing temporal and kinetic aspects of the stance phase of gait. The [Formula: see text]-nearest neighbors algorithm achieves the best machine learning classification accuracy of 95%. The preliminary research establishes the foundation for more sophisticated endeavors respective of the powered prosthesis, such as determining the appropriateness of modifying the software control architecture to best accommodate the progressive lifestyle evolutions and adaptations of the person with amputation.