Research Profile

 
On-going   On-hold   Funding

Recursive estimation of higher order statistics for hand prosthesis control
Royal Society
Details to come.

Co-adaptive Control for Myoelectric Prosthesis
The Leverhulme Trust

A myoelectric decoder would be greatly improved if it incorporates the capability to interact adaptively with the user’s motor system instead of relying on a fixed functional mapping between the myoelectric activity and the motor behaviour. Such interaction should be bi-directional exploiting the remarkable flexibility of the motor system to learn to operate novel devices. By providing the user with appropriate feedback, the prostheses should allow subjects to acquire an internal model of the new mapping between efferent command signals and motor output for accurate feed-forward control. Also, the decoder, informed by the user’s optimal pattern of muscle activity can converge to a mapping which cooperates maximally with the subject’s surviving motor architecture.

The last decade has seen radical advances in our theoretical understanding of optimality in natural motor control and these principles are beginning to inform the development of new human-machine interfaces. For instance, in pilot experiments, we trained subjects to control a computer cursor using non-intuitive mappings between muscle activity and cursor position. With practice, muscle-tuning functions became cosine-shaped and modulated so as to minimise the impact of signal-dependent motor noise on movement accuracy. However, these experimental methodologies have yet to be explored with respect to optimising the operation of myoelectric prostheses using altered musculature following amputation. In developing our novel co-adaptive design, we exploit the inherent trial-to-trial variability of repeated movements to characterise the constraints on optimising the muscle-prosthesis mapping. Provision of accurate proprioceptive feedback will facilitate subject’s acquisition of an internal model of this mapping. Finally, reinforcement learning provides the necessary theoretical framework for ensuring that such co-adaptive closed-loop synergistic processes converge at a stable solution.


Task-dependent structure of neuronal variability during abstract BMI control

Medical Research Council

Primates can volitionally modulate neuronal activity according to arbitrary reward rules and learn to control Brain-Machine Interfaces (BMIs) with randomized decoders. We have previously shown that in redundant myoelectric-control tasks, subjects improve accuracy by buffering trial-to-trial variability into task-irrelevant dimensions of the muscle space. However it is not known whether variability in the neural space can also be shaped in a task-dependent manner during control of a redundant BMI.

We therefore investigated the structure of neural noise covariation in two macaque monkeys performing a simple brain control task involving two neurons, C1 and C2. Over separate sessions, 81 pairs of neurons were sampled from chronically-implanted electrode arrays in M1 and PMv. The one dimensional position of a cursor was determined by the instantaneous firing rate of C1 alone or the sum (C1+C2) or difference (C1-C2) of the firing rates. The monkeys’ task was to hold the cursor for 500ms in targets that appeared in one of four different locations.

We calculated trial-to-trial variability of the firing rate trajectory in neural space. Trajectories for each trial were aligned to the first time that cursor entered the target. To remove any structure that this alignment imposed on the trajectory distribution, we characterised the evolution of neural variability relative to the point in neural space at which the cursor entered target. Variability was calculated along dimensions of positive (C1+C2) and negative (C1-C2) covariation and defined an index of covariation (IoC) to quantify correlation structure. A zero IoC indicated a uniform pattern of variability while the limiting values of  imply the variability is entirely constrained to axes of positive/negative covariation in the neural space. IoC was evaluated up to 500ms after the alignment point.

We found an effect of task condition on neural IoC. Overall, the distribution of IoC tended towards positive values, consistent with a general pattern of positive noise correlation between neurons. During the C1-C2 control block, IoC was increased further suggesting greatest neuronal variability in the dimension of positive correlation (the task-irrelevant dimension). In contrast, IoC was, on average, slightly negative in the C1+C2 control condition. IoC during control by C1 alone fell between that of C1-C2 and C1+C2 conditions. This modulation of IoC suggests that trial-to-trial variability was to some extent shaped to reflect the task-relevant and -irrelevant dimensions of the neural space.

It remains to be seen whether the task-dependent structure in neuronal variability emerges from divergence in feed-forward inputs or feed-back correction of errors. In either case, our results suggest that redundancy in abstract BMIs can be exploited to improve accuracy of control by buffering variability into task-irrelevant dimensions of the neural space.

Flexible Cortical Control of Task-Specific Motor Synergies
Medical Research Council

It has been proposed that the motor system reduces the dimensionality of movements by combining a small number of motor primitives that are constrained by neural circuitry. Alternatively, covariation of synergistic muscles may arise as the solution to an optimisation problem constrained by task requirements. These theories are difficult to distinguish when studying natural movements because neural circuits may have evolved to reflect biomechanical properties of the limbs they control. We therefore investigated patterns of muscle covariation using a myoelectric-controlled interface in which task requirements and biomechanics could be dissociated.

16 subjects made repeated movements of a myoelectric cursor, receiving scores that reflected the proportion of time they remained within targets. Cursor position was determined by smoothed, rectified electromyogram (EMG) from two muscles acting orthogonally in the task space. Elliptical target shapes imposed relevant and irrelevant dimensions within the task space, oriented such that movement variability should optimally be constrained along dimensions of either positive or negative EMG covariance. Muscle combinations included natural synergists (FDI-APB, ECR-FCR) and unnatural pairs (FDI-ADM, ECR-APB).

Performance with all muscle pairs improved over 4 blocks of 99 trials, although scores were lower for unnatural pairs. Trial-to-trial movement variability was modulated appropriately by target shape, particularly during later blocks. Multiple regression analysis of EMG covariance showed the significant main effect of target shape for all the pairs. For natural synergists, the interaction between target shape and block number was also significant. This suggests subjects could learn to modulate muscle covariation on-line to minimise errors in task-relevant dimensions, particularly using FDI-APB.

To explore the neural circuitry underlying this ability, we calculated intermuscular coherence between FDI and APB during hold periods. In later blocks, beta-band coherence was modulated by target shape, with maximal coherence when target orientations required positive covariance between muscles. This indicates that increased cortico-spinal drive from a common source mediates task-specific patterns of muscle covariation.

Our results show that rather than being limited to a small number of fixed synergies, the human hand can recruit a wide repertoire of co-ordinated muscle patterns appropriate for task demands. The divergence in cortico-spinal projections to distal motoneurons may provide a rich neural substrate for flexibly minimising movement errors in task-relevant dimensions.

EMG Prediction from Neural Recordings in the State-Space
The Wellcome Trust

We proposed a constrained point process filtering mechanism for prediction of electromyogram (EMG) signals from multi-channel neural spike recordings. Filters from the Kalman family are inherently sub-optimal in dealing with non-Gaussian observations, or a state evolution that deviates from the Markovian setting assumption. To address these limitations, we modeled the non-Gaussian neural spike train observations by using a generalized linear model (GLM) that encapsulates covariates of neural activity, including the neurons' own spiking history, concurrent ensemble activity, and extrinsic covariates (EMG signals). In order to predict the envelopes of EMGs, we reformulated the Kalman filter (KF) in an optimization framework and utilized a non-negativity constraint. This structure characterizes the non-linear correspondence between neural activity and EMG signals reasonably. The EMGs were recorded from twelve forearm and hand muscles of a behaving monkey during a grip-force task. For the case of limited training data, the constrained point process filter improved the prediction accuracy when compared to a conventional Wiener cascade filter (a linear causal filter followed by a static non-linearity). The approach was tested for different bin sizes and delays between input spikes and EMG output; 20 ms bin size and 40 ms delay provided the highest prediction rates. For longer training data sets, results of the proposed filter and that of the Wiener cascade filter were comparable.

Brain Computer Interfacing with steady-state movement related potentials
The Wellcome Trust

An approach for BCI in the STF domain was developed in which the tensor of the time-varying spectrum of the multi-channel EEGs was decomposed into an appropriate number of spatial, temporal, and spectral signatures. By STF decomposition of EEGs in the mu band (8-13 Hz), EEG dynamics during left and right index movements were investigated. The spatial signature of the identified movement related factor in each recording trial was used as a feature for classification. However, the computational complexity of the STF modeling limits the use of such BCI in real-time applications. After all, the question was “Do STF signal processing approaches consistently outperform the conventional time-space or time-frequency methods in brain signal analysis?” One might argue that although the STF models seem to be effective, their implementation on the brain signals is only justified when a phenomenon of interested sparsely occurs at least in one of those domains.

An alternative approach based on the steady-state movement related potentials (ssMRP) recorded during real finger movements was initiated. The backbone idea was to modify the EEG recording protocol in order that the subjects produce continuous spatio-spectrally localized motor related potentials. Therefore, meet the sparsity condition in the space and frequency domains are met. It was shown that ssMRPs provide a measure of sensorimotor cortex activation during rhythmic tapping and be a potential signal for a real-time high accuracy BCI. The main advantages of the ssMRP-based BCI over other approaches are its simple recording setup and straightforward computations.

We, in collaboration with Dr Jakub Stastny, further looked into the application of the hidden Markov models for classification of ssMRP [see here], this is an on-going work.


Artifact Removal from EEGs
The Leverhulme Trust

The EEG potentials can be severely contaminated by various artifacts such as eye-blinks (EB) and eye- or body- movements. Essentially in BCI applications, artifacts interfere with the processing algorithms and may cause unacceptable BCI outputs. During my PhD, two (conceptualy similar) methods for EB artifact removal from EEGs in the space-time-frequency (STF) domain were developed. First, an effective semi-blind signal extraction (SBSE) algorithm in which the spatial distributions of the EB artifacts were identified by using the STF modeling of contaminated EEGs. These signatures were then utilized in a BSE based removal stage. However, EEG decomposition to the STF signatures is computationally intensive and also the performance of the SBSE method depends on how precise the EB spatial signatures are estimated. In order to surpass these two limitations, a hybrid “STF modeling”-“robust minimum variance beamforming (RMVB)” framework was developed. In particular, an STF-TS model for EEGs for decomposition of the EB contaminated EEGs was proposed and it was shown that the STF-TS model closely approximates the classic STF model, with much lower computational cost. The spatial signatures were then exploited in the RMVB paradigm to extract and remove the artifact.

This work has been nicely extended by Dr Yodchanan Wongsawat.

 

Myoelectric Signal Classification for Prosthetics Control

We proposed an approach for surface electromyogram (sEMG) signal classification which utilizes higher order statistics of sEMG signal to classify four primitive motions, i.e., elbow flexion, elbow extension, forearm supination, and forearm pronation.The sEMG signal generated during isometric contraction is modeled by a stationary process whose probability density function is assumed to be either Gaussian or Laplacian. Using the negentropy measure, it was shown that the level of non-Gaussianity of sEMG signal recorded in muscular forces below 25% of maximum voluntary contraction (MVC) is noticeable. An accurate classification is achieved by using the sequential forward selection method for reducing of the dimensionality of feature space and the K-nearest neighbor (KNN) classifier. The results indicated that including higher order statistics of the sEMGs increases the correct classification rates.

Last Updated Feb 3, 2012