Enhancing the Control of Prosthetic Hands

Aims:
* Assess which confounding factors are most prevalent in electromyography (EMG) variability affecting performance in the abstract decoding prosthesis control scheme.
* Investigate learning and consolidation of a new motor task to a 'mental map' in humans. Subsequently assess the ability of flexible adaptation of the acquired map.
* Explore how simultaneous machine and user learning interact and develop with each other.
* Employ the following compensatory techniques to ensure more robust control.

Objectives:
* Measure the effect of different circumstances (i.e. confounding factor influence or co-adaptive learning) on user performance on the abstract control task, in terms of score, path efficiency, learning rate, and the Southampton Hand Assessment Procedure (SHAP) score.
* Develop adaptive algorithms, e.g. iterative solution manifold tracking, to adjust the prosthesis controller for the predicted effect of the identified artefacts. Key studies will include identifying the most efficient rate of adaption and its effect during daily use of the prosthesis.
* Test the developed platform with people with limb difference and evaluate the efficacy of developed platform in comparison to existing methods.

Overview: Abstract Decoders
Abstract decoding is a novel control scheme for motorised prosthetic hands, motivated by motor learning. The user is required to navigate a cursor in a two-dimensional task space towards a desired target which corresponds to a predefined movement type. The abstract decoding control scheme only requires a single pair of electrodes. Muscle activity at each site independently dictates the movement of a cursor along a single axis. Subjects have been shown to continually improve performance in terms of accuracy as the number of practice trials increases. In this method, the user learns to generate functional EMG patterns by contracting muscles that are not naturally used to control the grasp. This notion is completely in contrast to classical pattern recognition decoders in which the prosthesis learns to identify movement intent(s) by decoding the EMG patterns without considering the users' learning capability.

Confounding Factors
There are many factors that can contribute to the variability in the EMG signal resulting in a degradation of the user's performance whilst operating the prosthesis. It is unknown what effect certain factors such as a change in limb position or electrode shift have on abstract decoding. The common approach of mitigating these effects in other decoders often calls for applying computationally intensive algorithms on high density data, which typically results in prolonged classifier training. This is undesirable as the long training time may frustrate the user, increasing the likeliness of prosthesis abandonment. This general approach is also thought to not scale well under simultaneous multi-factor influences. The addition of extra sensors and the need for higher-end processing components results in greater cost and further increases the infeasibility of clinical funding.

The nature of abstract decoding is distinct to other control schemes. Since it relies on human motor learning there is no need for classifier training. In addition, it requires only two input channels to restore several hand postures and involves comparatively less-complex computations. Therefore, attempts to mitigate these confounding factors will take a distinct approach. As such, unique solutions are expected to arise.

Research Questions:
1. How do specific confounding factors affect user performance under abstract control?
2. To what extent can the user learn to compensate for perturbations in the 'mental model' with a decoder that accommodates learning?
3. What effect does machine adaptation have on user performance and learning?
4. 'When to adapt' and 'how to adapt' the control policy? - Both in terms of co-adaptive learning & compensatory adjustments to confounding factors