Evaluation of data mining approaches for the control of multifunctional arm prostheses

Abstract

Commercially available hand prostheses are controlled by myoelectric signals, biosignals that must be accurately interpreted if a prosthesis is to work effectively. We propose the use of standardized muscle contractions to select grasp types; these are taught to the prosthesis using data mining methods and parameters individually adapted from the patient. Embedded systems in medical devices often restrict the application of classification algorithms due to a lack of computational power and memory; therefore, important issues for the design of medical devices are not only classification accuracy, but also computational complexity and low assignment of memory. This paper introduces a new method for feature selection and examines techniques for feature dimension reduction, namely feature selection and aggregation. Wrapper approaches with modified measures were introduced to improve classification results; the effects of this approach are discussed using a synthetic benchmark. Our analysis of measured biosignals, obtained from the control of hand prosthesis by seven different subjects, demonstrates that the proposed approach works for real-world problems.