Personalized Fusion of Ultrasound and Electromyography-derived Neuromuscular Features Increases Prediction Accuracy of Ankle Moment during Plantarflexion

Objective: Compared to mechanical signals that are used for estimating human limb motion intention, non-invasive surface electromyography (sEMG) is a preferred signal in human-robotic systems. However, noise interference, crosstalk from adjacent muscle groups, and an inability to measure deeper muscle tissues are disadvantageous to sEMG’s reliable use. In this work, we hypothesize that a fusion between sEMG and in vivo ultrasound (US) imaging will result in more accurate detection of ankle movement intention.

Methods: Nine young able-bodied participants were included to volitionally perform isometric plantarflexion tasks with different fixed-end ankle postures, while the sEMG and US imaging data of plantarflexors were synchronously collected. We created three dominant feature sets, sole sEMG feature set, sole US feature set, and sEMG-US feature fusion set, to calibrate and validate a support vector machine regression model (SVR) and a feedforward neural network model (FFNN) with labeled net moment measurements.

Results: The results showed that, compared to the sole sEMG feature set, the sEMG-US fusion set reduced the average net moment prediction error by 35.7% (p<0.05), when using SVR, and by 21.5% (p<0.05), when using FFNN. In SVR, the sole US feature set reduced the prediction error by 24.9% (p<0.05) when compared to the sole sEMG feature set. In FFNN, the sEMG-US fusion set reduced the prediction error by 28.2% (p<0.05) when compared to the sole US feature set.

Conclusion: These findings indicate that the combination of sEMG signals and US imaging is a superior sensing modality for predicting human plantarflexion intention and can enable future clinical rehabilitation devices.