A training data reduction scheme for electromyography, force myography, and their combination under varying limb positions and loading conditions

Abstract

Devices that rely on hand gesture recognition as a form of control require accurate and reliable classification performance to ensure effective use. Two techniques used for this purpose include electromyography and force myography, which rely on consistent muscle activity for accurate hand gesture classification. However, their effectiveness is reduced by changes in limb position and loading during object interaction, which introduce variations in the recorded signal, thereby increasing the rate of gesture misclassifications. Since limb movement and object interactions are crucial aspects of how we use our hands in daily life, our previous work investigated how the changes in limb position and loading affected three sensing modalities: electromyography, force myography, and their combination. Here, we build upon our previous work and develop a strategic training strategy to reduce the requisite training data, and correspondingly, required time, while preserving gesture classification accuracy for these three muscle measurement modalities. We used a forward wrapper method to quantify the impact of reducing training data, while also identifying a threshold where additional training data no longer benefits classification accuracy. These results can be used to strategically select training conditions for hand gesture classification systems to reduce training time and ensure reliable classification.