Functionally Adaptive Myosite Selection Using High-Density sEMG for Upper Limb Myoelectric Prostheses

Abstract

Objective: Our study defines a novel electrode placement method called Functionally Adaptive Myosite Selection (FAMS), as a tool for rapid and effective electrode placement during prosthesis fitting. We demonstrate a method for determining electrode placement that is adaptable towards individual patient anatomy and desired functional outcomes, agnostic to the type of classification model used, and provides insight into expected classifier performance without training multiple models.

Methods: FAMS relies on a separability metric to rapidly predict classifier performance during prosthesis fitting.

Results: The results show a predictable relationship between the FAMS metric and classifier accuracy (3.45%SE), allowing estimation of control performance with any given set of electrodes. Electrode configurations selected using the FAMS metric show improved control performance ( ) for target electrode counts compared to established methods when using an ANN classifier, and equivalent performance ( R² ≥ .96) to previous top-performing methods on an LDA classifier, with faster convergence ( ). We used the FAMS method to determine electrode placement for two amputee subjects by using the heuristic to search through possible sets, and checking for saturation in performance vs electrode count. The resulting configurations that averaged 95.8% of the highest possible classification performance using a mean 25 number of electrodes (19.5% of the available sites).

Significance: FAMS can be used to rapidly approximate the tradeoffs between increased electrode count and classifier performance, a useful tool during prosthesis fitting.

Fig. 1.

Examples of electrode locations and resulting class separation that might result from various fitting methods. The separation plots illustrate EMG datapoints from two motion classes after dimensionality reduction. Increased overlap between class data clusters signifies more difficult user intention estimation. (a) Conventional electrode configuration - untargeted electrode placement for a trans-radial subject. (b) Myosite selection using palpation method for targeted electrode placement - high intensity locations recognized by muscle movement are likely to get selected using this method. (c) The FAMS method chooses electrode locations to maximize separation between movement classes, and allows prosthetists to make informed decisions about tradeoffs between expected performance and electrode count at the time of fitting.

Fig. 2.

The FAMS algorithm for optimal sEMG channel ranking simultaneously maximizes inter-class separation of data while minimizing interchannel information redundancy. (a) An illustrative example of a signal that might be produced when the subject is asked to perform three different motion classes. Windowed Mean Absolute Value (MAV) features are extracted from the recorded training data for each channel (see Section II-B.3). The inter-class distances are used to create multi-dimensional models of the signals generated by each class. Every myosite used corresponds to a dimension of the model. (b) Models formed for each class using data from a single electrode channel. The probability of confusion between any two classes is related to the overlap between the Gaussian models. Assuming equal priors for all classes, the likelihood of misclassification for any given class $U$ can be bounded by summing the pairwise Bhattacharyya coefficients between $U$ and all other classes. This bound is less accurate for low dimensional classes, when the volume of multi-class intersections is nontrivial. (c) A 2D model formed with data from two electrode channels. As the number of channels increases, the volume of multi-class intersections diminishes rapidly, leading to a more accurate error bound. The expected distance $D$ (Eq. 6) can be used as a heuristic for expected accuracy.

Fig. 3.

A histogram shows that features extracted from a single myosite electrode can be accurately modeled using 1-dimensional Gaussian models to represent each class. This particular channel produces very good separation between the ‘rest’ (RE) and ‘hand open’ (HO) classes, but no separation between HO and ‘hand close’ (HC). Thus to make accurate classification possible, it is necessary to include other myositis that introduce separation between HO and HC.

Fig. 4.

(a) Custom HD electrode array showing all 128 electrode contacts. (b) 128 channel electrode array wrapped around the subject’s residual limb.

Fig. 5.

Aggregate classification accuracy (Able-bodied + Amputees) comparison using (a,b,c) TCN and (d,e,f) LDA classifiers for different myosite configurations. Myosites were selected using the FAMS method, LASSO, DEFS, SFS, and compared to a typical 8EQUI configuration. This comparison was done for (a,d) 5 movement classes (RE HD, HC, PR, SU), (b,e) 7 movement classes (RE HD, HC, PR, SU, WF, WE) and (c,f) 11 movement classes (RE HD, HC, PR, SU, WF, WE, IP, KE, TR, PI).

Fig. 6.

The minimum number of electrodes to saturate classification performance (N-OPT) was calculated for each subject using different channel selection methods. N-OPT distributions are plotted for the various classification techniques. With the LDA classifier, there is no statistically significant difference in saturation point between any selection methods, although FAMS shows faster saturation for low class counts. FAMS consistently outperforms LASSO with a TCN classifier and MAV feature set.

Fig. 7.

TCN models were trained and evaluated using various myosite configurations for each subject. Here the classification accuracy for those models is compared against the normalized distance metric $D$ for various electrode configurations and functional movement sets. A second-order line fitted using Weight Least Squares shows an R² value of 0.685. From this we can conclude that $D$ is a reasonable heuristic for classification accuracy.

Fig. 8.

Myosite configurations generated using the FAMS selection method for 5,7, and 11 movement classes. Myosites are presented in order of addition to the set. Lighter sites were added sooner than darker ones, and have a higher impact on classification performance. The number of electrodes $N$ selected varies with the movement set as well as the subject under consideration. For example, for Subject Amputee 1 , N={20,28,32} for 5,7 and 11 class movements respectively, whereas for Subject Amputee 2 N={40,44,48} for 5,7 and 11 class movements respectively. Although some of the electrode locations are common between configurations, there is notable variation between the overall placements and which myosites are most important. This highlights the need for targeted electrode placement techniques that adapt to individual patient needs.

Fig. 9.

128 channel HD sEMG arrays showing myosites from different selection methods and anatomical correspondence of selected myosites: a) 8-EQUI configuration- all 8 myosites have equal significance. b) Muscle activity map showing sites likely to be selected by palpation: high intensity sites (in red) are most likely to be selected. c) HD sEMG arrays showing 8-OPT myosites selected by FAMS system and forearm with corresponding SIM showing myosites most likely to be selected. Principal component (PC) projection show that data from the first 8-OPT myosites creates much better class separation for grip classes than data from 8-EQUI configuration. (d) PC Projection for data from 8-EQUI(untargeted) configuration. (e) PC projection for data from myosites likely to be selected during by palpation. (f) PC projection for data from 8-OPT myosite configuration.