Predicting biological joint moment during multiple ambulation tasks.

Combining machine learning models with wearable sensing provides a key technique for understanding the biological effort, creating an alternative to inverse dynamics based on motion capture. In this study, we demonstrate a novel approach to not only estimate but predict the joint moment in advance for multiple ambulation modes. By combining electromyography (EMG), inertial measurement units (IMU), and electrogoniometers, we enable the prediction of the joint moment only from wearable sensors. We performed a forward feature selection to determine the best feature sets for different anticipation times of the intended moment generated at the hip, knee, and ankle, encompassing level walking on a treadmill and ascent/descent of stairs and ramps. We show that wearable sensors can predict the joint moment with an MAE of 0.06 ± 0.02 Nm/kg for direct estimation and an MAE of 0.10 ± 0.04 Nm/kg when predicting 150 ms in advance, corresponding to an MAE within 9.2% of the joint moment range. We found that the hip moment had a significantly lower error than the knee and ankle when anticipating the joint moment (Bonferroni test, p < 0.05). The accurate estimation of the joint moment could monitor user activity to reduce risk factors and inform the control of exoskeletons.