Continuous grip force estimation from surface electromyography using generalized regression neural network

Abstract

BACKGROUND:

Grip force estimation is highly required in realizing flexible and accurate prosthetic control.

OBJECTIVE:

This study presents a method to accurately estimate continuous grip force from surface electromyography (sEMG) under three forearm postures for unilateral amputees.

METHODS:

Ten able-bodied subjects and a transradial amputee were recruited. sEMG signals were recorded from six forearm muscles on the dominant side of each able-bodied subject and the stump of amputee. Meanwhile, grip force was synchronously measured from the ipsilateral hands of able-bodied subjects and contralateral hand of amputee. Three force profiles (triangle, trapezoid, and fast triangle) were tested under three forearm postures (supination, neutral and pronation). Two algorithms (Generalized Regression Neural Network (GRNN) and Multilinear Regression Model (MLR)) were compared using several EMG features. The estimation performance was evaluated by coefficient of determination ( $R^2$ ) and mean absolute error (MAE).

RESULTS:

The optimal regressor combining TD and GRNN achieved $R^2=$ 96.33 $\pm$ 1.13% and MAE $=$ 2.11 $\pm$ 0.52% for the intact subjects, and $R^2=$ 86.86% and MAE $=$ 2.13% for the amputee. The results indicated that multiple grip force curves under three forearm postures could be accurately estimated for unilateral amputees using mirrored bilateral training.

CONCLUSIONS:

The proposed method has the potential for precise force control of prosthetic hands.