Comparing the Outcomes of Population-averaged and Personalised Input Feature Selection for Transhumeral Prosthetic Interfaces

Abstract

An interface for a powered upper limb prosthesis is designed to achieve the target poses intended by the human user as interpreted from input features measured from a variety of sensors typically worn on the residual limb. Selection of which input features to use in the interface is usually done based on which features provide the best likelihood to identify the intended target poses. This work investigates the outcomes of the prosthetic interfaces constructed with input features selected based on the performance of all subjects (population-averaged) vs. the performance of individual subjects (personalised). Specifically, two outcomes are evaluated. First, the diversity in the resulting input feature sets across different subjects is evaluated. Second, the accuracy in identifying the intended target pose using the input features selected through personalised and population-averaged data are analysed and compared. The experiment was conducted with 10 able-bodied subjects in a virtual reality (VR) platform. The corresponding kinematic information of their arm movements and muscle activities were recorded and utilised to construct the input features in the scenario of transhumeral prostheses. The outcomes show a significant variation in the resulting input feature sets from subject to subject. The accuracy of the personalised feature sets is found to be higher in classifying the intended target poses. For the case of selecting 2 out of the 41 input features constructed from the recorded sensor data, none of the 10 subjects shares any common set of input features and the accuracy is improved by 10.4±6.8% with the personalised selection. The outcomes demonstrate the extent of person-to-person variation that needs to be taken into account when designing prosthetic interfaces and the quantified potential gain (albeit in a specific application) from personalising interfaces.