MMG-torque estimation under dynamic contractions

Torque estimation using mechanomyographic (MMG) signal is typically calculated by the root mean square (RMS) amplitude. Raw MMG signal is processed by rectification, low-pass filtering, and mapping to estimate torque. However, one-to-one mapping is not accurate because if the input is interfered by noise, the output follows directly. In this work, beside RMS amplitude, another significant feature of MMG signal, i.e., frequency variance, was found and used for constructing the MMG-torque estimator. Seven subjects produced constant posture and torque contractions about the elbow while MMG signal and torque were recorded. We found that MMG RMS amplitude increased monotonously and frequency variance decreased under incremental voluntary contractions. A MMG-torque estimator was introduced using MMG RMS amplitude and frequency variance as inputs and a two-layer neural network as the modeling algorithm. Experimental evaluation of the estimator was done under constant posture and sinusoidal contractions at 0.5Hz, 0.25Hz, 0.125Hz, and random frequency. The results of the proposed MMG-torque estimator and MMG RMS amplitude linear mapping were also compared. The estimation of MMG-torque estimator has better accuracy than linear mapping for all contraction frequencies. The mean absolute error decreased 6% for the 0.5Hz contraction, 43% for 0.25Hz contraction, 52% for 0.125Hz contraction, and 30% for random frequency contraction.