Machine-Learning-Based Accurate Finger Joint Stiffness Estimation With Joint Modular Soft Actuators

Abstract

Finger joint stiffness assessment is very important in quantifying the degree of disability in stroke patients and determining rehabilitation strategies and plans. However, the evaluation methods used in clinical practice rely heavily on the experience of clinicians. In our previous study, we adapted an analytical model originally proposed for a whole-finger soft actuator to objectively quantify joint stiffness by using a new type of joint modular soft actuator designed for individualized hand rehabilitation. However, stiffness could not be accurately estimated owing to the effects of the interaction between joint modular soft actuators and a finger, as well as, between the actuators themselves, which could not be represented by the adapted analytical model. In this study, artificial neural network (ANN)-based models were proposed to simultaneously quantify the stiffness of three joints using joint modular soft actuators and compared with the adapted analytical model and other machine-learning (ML)-based models. Moreover, the estimation performance was verified for high stiffness values and different finger sizes. The results show that the ANN-based models estimate stiffness more stably and accurately than the adapted analytical and other ML-based models. This study shows the feasibility of quantitative evaluation of joint stiffness using joint modular soft actuators.