Information theoretical measures from ultrasound data for human motion understanding

Abstract

Noninvasive ultrasound (US) sensing has been recently introduced as an intuitive human-machine interface. Most research to date has focused on using US images of muscle to predict human movement intentions. However, the richness of unprocessed US signals as a source of neuromuscular information may have been left unnoticed. In the present study, we investigate the feasibility of using unprocessed US signals of muscle tissue to continuously predict knee motion kinematics during seated knee flexion/extension and sit-to-stand movements. Unprocessed US signals were compared to US images using a mutual information analysis to quantify the information gained from each of these US data forms about knee motion. The motion prediction accuracy of unprocessed US signals was compared to motion prediction accuracy of US images. It was observed that a statistically comparable amount of information can be gained from both US data forms $(p\lt 0.05)$. The prediction accuracies were also statistically comparable $(p\lt 0.05)$, and average root mean squared error for knee angle prediction was 1.66° when using unprocessed US signals compared to 2.25° when using US images. Noteworthy, the computation speed was around 33 frames per second (FPS) when using US images compared to 251 FPS when using unprocessed US signals. Overall, this study highlights the promise of unprocessed US signals as a source of neuromuscular information for human motion prediction in real-time, while omitting the signal processing steps required to reconstruct the US images and the associated engineering sophistication, facilitating the future integration of US sensing as a human-machine interface.