Context- and Physiology-aware Machine Learning for Upper-Limb Myocontrol
by Gauravkumar K. Patel
Date of Examination:2018-05-03
Date of issue:2018-05-16
Advisor:Prof. Dr. Dr. Dario Farina
Referee:Prof. Dr. Dario Farina
Referee:Prof. Dr. Marcus Baum
Files in this item
Name:Patel PhD Thesis.pdf
Size:4.86Mb
Format:PDF
Description:Thesis PDF file
Abstract
English
The world around us is shaped in such a way that our hands are necessary to accomplish most activities of daily living. It is therefore undeniable that the loss of the upper limb, partial or total, represents a severe impairment. With current advancements in robotic technology, it is now possible to replace a missing limb with a dexterous upper-limb prosthesis. However, the development of a reliable human machine interface (HMI), connecting the user and the prosthesis, is still an open challenge. Essentially, the HMI defines an invariant mapping scheme to transform electromyogram (EMG) signals generated by the user into movements on the prosthetic device, thereby allowing the user to control available functions by generating appropriate (predefined) EMG signals. An HMI control driven by EMG signals is known as myoelectric control or myocontrol. EMG signals associated with a particular motor task are distinct and repeatable and therefore, it is possible to use one of the many well-known machine learning (ML) algorithms as HMI for estimating different user motor intentions. With ML-based HMIs, users can directly activate a desired prosthesis function by producing EMG signals associated to that function during supervised learning. Although conceptually promising, ML-based control has shown a limited clinical viability, mainly due to the lack of reliability and robustness during real time use. The aim of this thesis was to improve the reliability and robustness of ML based control by developing context- and physiology- aware ML methods for upper limb myocontrol. Today, most ML methods used for myoelectric control follow the conventional pattern recognition paradigm, where training data is collected using a supervised procedure and a mathematical function is fitted over the collected data to define an invariant mapping scheme between the user’s EMG and available prosthesis movements. This conventional approach has two limitations. First, the mapping scheme (between the EMG and available movements) remains static (invariant) during use and does not consider the dynamics associated with real-life use of prosthesis. Second, the mathematical function fitted over the training data is assumed to implicitly capture the physiological principles behind generation of EMG; this assumption might not be true, as many commonly applied ML methods do not model the underlying physiology. The first limitation can be solved by developing ML methods which can consider context information describing the state of the system and/or environment during prosthesis use. This context information can be acquired either directly from the user or by placing additional sensors (e.g. inertial units) on the prosthesis. The former idea of deriving context information from the user is quite interesting, as it gives to the ML an opportunity to improve control by considering user’s requirement(s) during use. This thesis proposes one ML method (called Modular Regression, see Chapter 2) which exploits user-generated context information to improve control for different activities of daily living (ADL). Specifically, the proposed ML method organizes each prosthesis function as a module, which the users can insert/remove as required to best accomplish a given ADL. Next, if additional sensors were placed on the prosthesis to automatically derive context information, the ML controller would get an opportunity to (automatically) monitor the state of the prosthesis and react accordingly to maximize reliability and robustness. This thesis proposes one ML approach (called context driven control, see Chapter 3) which utilizes context information from additional sensors to model different prosthesis states and then, the parameters of ML control were adapted to mitigate expected disturbances in each prosthesis state. Thus, with both new ML methods, the mapping scheme (between the user’s EMG and available movements) does not remain static, but becomes reactive to the context information coming from the user or additional sensors. Experiments involving functional tasks were conducted to compare the newly developed context-aware ML methods with the conventional ML-based control. The experimental results indicate that the context-aware methods significantly outperform conventional ML control. The second limitation of conventional ML approaches, i.e. the fitted mathematical function may or may not capture the latent physiology information, can be solved by designing ML methods that are aware of the underlying muscle physiology. This thesis presents one ML algorithm (based on the cosine similarity metric, see Chapter 4) which exploits the principle of muscle coordination to classify EMG for online myoelectric control. Specifically, the principle of muscle coordination states that force production for a given movement relies on the coordination of different muscles and the EMG amplitude of involved muscles scales uniformly with the amount of force exerted. And therefore, the presented physiology-aware ML method was designed based on the assumption that amplitude-related EMG features for each movement are distributed along the line joining the origin of the feature space and the average maximum voluntary contraction of the movement. This assumption led to a simple training procedure and a computationally efficient solution. The presented physiology-aware ML method was extensively compared with the state-of-the-art ML method using four functional tasks. The results indicated that the new method performs significantly better than the standard ML method, while utilizing less training data and smaller computational effort. Overall, this thesis points to the potential advantage(s) of ML methods that exploit context and physiology information for online myocontrol over standard ML methods (with a static mapping scheme and no modelling of physiology), which largely prevail in the literature. Moreover, all ML methods presented in this thesis are simple, robust and computationally efficient, and therefore, they can be directly used for interfacing most prosthetic devices available in the market, with a minor hardware upgrade.
Keywords: machine learning, myocontrol, EMG, context-awareness, sensor data fusion, user-interaction