Ph.D. Defence - Muhammad Asim Waris
Muhammad Asim Waris will defend his Ph.D. thesis: "Multi-day analysis of surface and intramuscular EMG for prosthetic control".
26.06.2019 kl. 13.00 - 16.00
Myoelectric control of upper limb prostheses has been an active area of research for many years. To date, surface electromyography (sEMG) is the main control signal used in commercial systems. However, it has been suggested that signals obtained from implantable electrodes, such as intramuscular EMG, may be a better source to provide independent sources for control. From the control schemes point of view, pattern recognition (PR) has been extensively researched as means to enable a more robust, intuitive, effective and simultaneous control of a large number of degrees-of-freedom (DoF) as offered by current advanced prosthetic limbs. However, how to achieve robustness over time with different PR schemes has received less attention in the literature. In the current thesis, the main aim was therefore to examine the behaviour of EMG based PR myoelectric control over time. Three specific research questions (SRQ) were formulated to address this aim: 1) To what extent threshold values affect the time domain features and their combinations in surface and intramuscular recordings? This question was addressed in the study I, where the threshold for each feature was computed as a factor (R = 0:0.02:6) times the average root mean square of the baseline. For each threshold value, classification error was quantified using two classifiers first for each individual feature and then combined. Results have demonstrated that using appropriate threshold value is very important to assure acceptable performance. 2) What is the correlation between the performance of PR based myoelectric control schemes and time? This question was addressed in study II and III using surface and intramuscular EMG concurrently recorded from 10 able-bodied subjects and six trans-radial amputees for seven consecutive days. A standard linear regression analysis was performed in study II on each EMG type for the identification of time effect (days) on classification accuracies. Study II showed that performance is significantly dependent on the time elapsed between training and test. In study III, Artificial neural network outperformed all other tested classifiers in terms of mitigating the effect of time on classification. 3) How do PR training strategies influence real-time performance over time? This question was addressed in Study IV, an experimental protocol was designed to determine the effect of training strategies on real-time PR control over time using a Fitts’ law approach. Results suggest that increasing the size of training set over time (by concatenation) can be beneficial to assure robust performance of the system over time. Moreover, classification error can be mitigated as the time lag between training and testing increase.