• EMG
• control
• prosthesis
• electromyography

This thesis describes the efforts done to research on two closely related subjects regarding prosthetic applications. In the first part, this work treats the problem of onset detection using the electromyographic (EMG) signal during hand and wrist movement, typically achieved through the steady-state of the EMG. To the reach of out knowledge, little research on exploiting the transient of the signal has been conducted. Starting from a state-of-the-art onset detector (S1) based on the transient signal, a new strategy (S2) was designed and implemented. A different signal preprocessing pipeline was used and new criteria for threshold estimation were defined. The new strategy was validated and compared against the state-of-the-art using an existing dataset of combined wrist and hand movements. Results showed a significantly higher F1-score for the problem of onset detection for S2. Further analysis also showed a greater time consistency for S2, namely the property of detecting the onsets always at the same spot along the uprising curve. Extrinsic validation, via classification of the detected contractions, produced higher accuracies for S2.

The second section of this study approaches the problem of fine controlling a prosthesis using the EMG signal under a regression paradigm. To investigate this, we first designed an experiment in which subjects where required to match the force exerted by a precision grasp to a time-varying trace. The traditional method of regression using the root mean squared (RMS) value of the EMG amplitude was compared to a recently developed technique based on estimating the motor unit spike trains (MUSTs) that gave rise to the EMG. These MUSTs are used to compute the population firing rate, which is in turn used in the regression task. After conducting the experiment, results showed that when using only low-frequency contents (i.e. up to 2 Hz) of the firing rate, the error during tracking is comparable to that associated with the RMS approach, regardless of the number of motor units (MUs) estimated. If higher frequencies are included (i.e. up to 5 Hz) the performance of the former approach decreases, along with its stability. However, model predictions indicate that if more MUs were included in the process, the novel approach would outperform RMS-based regression.