Proceedings of the 2005 IEEE 9th International Conference on Rehabilitation Robotics June 28 - July 1, 2005, Chicago, IL, USA

WeB01-06

Fuzzy C-Means Clustering Analysis of the EMG Patterns of Six Major Hand Grasps
A. Bolu Ajiboye, MS and Richard F. ff. Weir, PhD, Member, IEEE

Abstract We seek to quantify the electromyographic similarities and differences between six major grasping patterns of the human hand: cylindrical, hook, palmar, lateral, spherical, and tip. Fuzzy c-means clustering (FCM) is a data reduction technique that allows continuous membership, on a [0,1] interval, of data points in the representative clusters. We found no grasp to be completely distinct from any other grasp, and found significant overlap of characteristics between several grasps. This overlap is quantified and discussed.

I. INTRODUCTION

HE human wrist and hand is an amazingly complex and versatile system with 27 bones and over 30 extrinsic and intrinsic muscles [2], resulting in 22 degrees of freedom [3]. Much of this versatility can be seen in the hands dexterity. Bernsteins definitions of dexterity include a coordinative structure capable of solving a particular motor task, a high adaptive power and flexibility at all levels of coordination, and the use of memorized corrective mechanisms [4]. Wiesendanger made this definition more concise to read, dexterity is the essence of purposeful motor behavior[entailing] all the ingredients of goal- and objectoriented actions, particularly manipulations that have been practiced and acquired [5]. This goal and object-oriented action can be seen in the typical grasp patterns that the hand can form. Napier classified all static hand grasps into two general categories: precision and power [6]. Precision grips were characterized by object contact primarily using the fingertips, whereas power grips were characterized by object contact primarily using the palm [7]. Other investigators have increased the grasp categories to include tripod, lateral, and tip prehensile grips, among others [8]. These categories are subdivisions of the larger power and precision categories. Of interest is the early and extensive work of Keller et al., who defined seven major categories of static hand posture for use in prosthetic devices. These are palmar 3-finger, palmar 2-finger, tip, lateral, hook, spherical, and cylindrical [1], as shown in Figure 1. According to Keller et al., the palmar 3finger grasp is the most common prehension pattern for

static grasping, while the lateral grasp is the most common in dynamic grasping. Current trans-radial single degree-offreedom prostheses mostly attempt to recreate the palmar 3finger prehension pattern [9]. It is readily observed that the grasps used for everyday objects may not fall explicitly into only one of these categories, but rather sometimes may be some combination of these grasps. The prehension patterns seem to share many characteristics kinematically, and may not be entirely distinct, as the categorizations suggest. This sharing of characteristics may also be evident through examination of the EMG patterns associated with each grasp. To what extent these grasps overlap has not been quantified. Current literature has used clustering algorithms to investigate mapping of individual joint and muscular activity in the M1 motor cortex [10, 11]. We have used fuzzy c-means (FCM) clustering of the EMG patterns of these grasps, in a tenmuscle parameter space, to quantify the overlap between these categories of hand grasps.

Figure 1. Keller et al. [1] defined seven major categories of prehensile grasps: a1) palmar 3-finger (3-jaw chuck) thumb opposes index and middle fingers ; a2) palmar 2-finger; b) tip thumb opposes index only; c) lateral; d) hook; e) spherical; f) cylindrical

II. METHODS Four subjects have thus far participated in the study. The pool consists of two men and two women, with a mean age of twenty-six years. All subjects had right hand dominance, with the exception of one woman. All subjects gave informed consent to the procedures as approved by the Northwestern University Office for the Protection of Research Subjects (NUOPRS) Institutional Review Board. Muscular characterization of the investigated hand postures was limited to ten extrinsic hand muscles: flexor digitorum superficialis (FDS), flexor carpi ulnaris (FCU), extensor carpi ulnaris (ECU), extensor digitorum communicus (EDC), pronator teres (PT), flexor pollicis longus (FPL), supinator, extensor pollicis longus (EPL), flexor carpi radialis (FCR),

Manuscript received February 14, 2005. This work was supported in part by the National Institute of Health (NIBIB/NICHD) under Grant 1 R01 EB01672-01. A.B. Ajiboye is with the Northwestern University Prosthetics Research Laboratory, Chicago, IL 60611 USA (phone: 312-238-6528; fax: 312-2386510; e-mail: [email protected]). R.F. Weir is with the Northwestern University Prosthetics Research Laboratory, Chicago, IL 60611 USA and the Jesse Brown VA Medical Center, Department of Veterans Affairs, Chicago, IL (phone: 312-2386512; fax: 312-238-6510; e-mail: [email protected]).

0-7803-9003-2/05/$20.00 2005 IEEE

49

and extensor carpi radialis (ECR). Even given that both intrinsic and extrinsic muscles participate in positioning and controlling the hand and wrist, the limitation placed on what muscles to record from were for two major reasons. First, subject comfort was of great importance. We wanted to minimize subject pain caused by the large number of needle sticks required by the protocol. Each muscle required two needle sticks for bipolar recording, resulting in a minimum requirement of twenty successful fine-wire placements. The subjects would not tolerate much more. Second, we ultimately seek to apply the acquired knowledge to the control of trans-radial myoelectric prostheses. In such an application, the intrinsic muscles are of no consequence. Muscles in the right forearm were first located by palpation and then with a standard tungsten probing electrode in accordance with instructions from standard electromyography text [12]. Two twenty-seven gauge monopolar needle electrodes were then inserted along the direction of the muscle, one 6.5 mm to either side of the monopolar probe, for a total bipolar separation of 13 mm. This was done for all ten muscles. The insertions were verified by having the subject perform tasks designed to isolate the muscle in question, and visually inspecting the simultaneously recorded EMG signals. Electrode insertion took on average four hours. Surface data was also simultaneously recorded with the intramuscular data, but that data is not considered here. The surface electrodes did not interfere with the intramuscular electrodes. All data was recorded with a Noraxon (Phoenix, AZ) Telemyo 2400R System, with a bandwidth of 20 1000 Hz, a gain of 2000, and an internal sampling rate of 2000 Hz. Before beginning the protocol, ten seconds of quiescent EMG was simultaneously recorded from the selected intramuscular control sites while the subject was relaxed, with the forearm supported and in a pronated position. Subjects were then instructed to perform each of six mimed grasp patterns: cylindrical, spherical, hook, palmar, lateral, and tip. Each grasp was first demonstrated for the subject, and then the subject practiced miming the grasp pattern until he/she felt comfortable. Upon a verbal cue, the subject, beginning from the relaxed position, mimed the instructed grasp, held it in steady-state until a verbal cue was given to stop, and then returned to the relaxed position. The total contraction time was about five seconds from cue to cue. Time was given for the subject to completely relax before the next trial of the same grasp commenced. The subject performed ten trials of each mimed grasp. Data recording began prior to giving the verbal cue to the subject and ended one to two seconds after the subject returned to rest. Raw EMG data from each trial was saved to a separate file. All files were processed using an in-house script written in Mathworks MATLAB (Natik, MA). Raw data was filtered between 10 300 Hz, full-wave rectified, and then smoothed with a 20 Hz filter. Contraction onset for each trial was determined as the first sample in a set of

consecutive samples that were five standard deviations above the quiescent mean. The quiescent mean of each signal was determined from the first second of data recorded prior to the initiation cue being given to the subject. All signals were then normalized to the respective maximum value observed throughout the protocol. We used fuzzy c-means (FCM) clustering [13, 14] to assess the extent to which the grasping patterns overlapped electromyographically. FCM seeks to group sampled data together so as to minimize the variance between data in the same cluster and maximize the variance between data in different clusters. This reduction method allows each cluster of data to be represented by a cluster center, where each center is a representation of the data geometrically closest to it. All data sets can belong to all cluster centers, with a degree-of-membership (DOM) in each cluster in the interval [0,1]. The DOM is directly related to the Euclidean distance between each data sample and the cluster center. Data samples and cluster centers are r length vectors, where r is the total number of EMG inputs. Equations (1)-(3) represent the restrictions on the clusters. 1) Every data sample must belong to a cluster, 2) the membership of each data sample in the universe of clusters must sum to one, and 3) no single cluster is equivalent with the null set or the universal set (i.e., c, the number of clusters, is greater than one and less than n, the number of data samples).

UA = X
i i =1

(1)

i =1

ik n

=1

(2)

0 < ik < n
k =1

(3)
c 2 m'

J m (U , v) = (ik ) (dik )
k =1 i =1

(4)

A matrix U is defined to express the membership of each sample (denoted by columns) in each cluster (denoted by rows), where membership of the kth sample in the ith cluster is denoted ik on the inclusive interval [0,1]. Given a vector of cluster centers v, an objective function (4) is defined, where dik is the Euclidean distance between the kth data sample and the ith cluster center. Minimization of equation (4) is performed through an iterative process implemented in standard mathematical software. The results are a final membership matrix U* that denotes the membership of each data sample in each cluster, and a final cluster matrix V* of dimension c x r, where c is the number of cluster centers and r is the number of inputs. U* is comprised of the DOMs of each data sample in each cluster group that minimizes the Euclidean distance between samples belonging to the same cluster and maximizes Euclidean distances between the

50

cluster centers themselves [14]. V* is comprised of the locations of the cluster centers in parameter space. Each trial is represented by a ten-element vector, where each element corresponds to the mean value of one EMG input during the transient period (first sec after onset) of the mimed grasp. These sixty data samples were then clustered by FCM. The number of clusters was specified to six, where each cluster represented a grasping pattern. The resulting U* matrix represented the membership of each trial in each grasp cluster. The same analysis was also performed separately using the steady-state portion of the mimed grasp (1 sec of data, beginning from sec after onset). III. RESULTS Figure 2(a) shows the U* membership matrix for one subject. Each rectangle represents the membership of one trial (columns) in a grasp cluster center (rows). Membership of each trial in each grasp cluster is on the interval [0,1]. If the grasps were completely different and did not share any properties electromyographically, Figure 2(a) would be a purely burgundy diagonal with a purely dark blue background. Figure 2(b) shows radial plots of the grasp memberships, averaging the results of the trials of each attempted grasp. So for example, this subject performed a hook grasp (top middle) that had a membership in the HOO (hook) cluster center of 0.723, in the LAT (lateral) cluster center of 0.108, in the PAL (palmar) cluster center of 0.063, in the TIP cluster center of 0.059, in the CYL (cylindrical) cluster center of 0.03, and in the SPH (spherical) cluster center of 0.017. It can be seen that, for this subject, the cylindrical grasp (self-membership of 0.902) was the most distinct electromyographically, followed by palmar (0.864), lateral (0.860), hook (0.723), spherical (0.718), and tip (0.654). Figure 2(c) shows the averaged results for all four subjects in the same format as Figure 2(a). Each grasp cluster is represented by a row and each subjects averaged mimed grasp membership is represented by a column. Table I(a) reports the average mimed grasp membership in each grasp cluster, based upon the transient portion of the mimed grasp, for the entire subject pool. On average, the tip grasp is the least distinctive electromyographically, with shared significance with the palmar grasp. This shared significance is reciprocated by the palmar grasp in the tip grasp cluster. Also of significance is the membership of the lateral mimed grasp in the palmar grasp cluster. The reciprocal is not as clear as in the palmar / tip relationship. Table I(b) reports the same information for the steady-state portion of the mimed grasp. Much of the distinctive nature of the grasps can also be seen in the steady-state. The rank of distinctiveness slightly changes however in that the hook grasp is now the most distinctive electromyographically.

(a)

(b)

(c) Figure 2. a) U* membership matrix resulting from FCM clustering of mimed grasp trials for a single subject. Each rectangle represents the membership of one trial (column) in each of the six grasp clusters (rows). Completely distinct grasp patterns would be characterized by a purely burgundy diagonal with a purely dark blue background. b) Polar plots showing the average membership of all trials in the grasp clusters. c) Average FCM clustering result for all four subjects.

IV. DISCUSSION Fuzzy c-means (FCM) clustering is one of many clustering algorithms, along with others like hard c-means (HCM) [14] and k-means, that attempt to reduce the amount of data to a representative set, in this case, based upon the Euclidean distances between the parameter space data points. All of these work to minimize the intra-cluster variance and maximize the inter-cluster variance. These other cluster methods, however, partition the data into non-overlapping clusters with hard boundaries. While these algorithms give

51

TABLE I(A) AVERAGE GRASP MEMBERSHIP FOR SUBJECT POOL (TRANSIENT) MIMED GRASP CYL CYL 0.717 0.063 0.044 0.057 0.060 0.058 HOOK 0.036 0.698 0.071 0.067 0.042 0.087 LAT 0.024 0.043 0.752 0.106 0.017 0.059 PAL 0.029 0.068 0.074 0.653 0.023 0.152 SPH 0.077 0.070 0.037 0.050 0.706 0.061 TIP 0.042 0.075 0.053 0.167 0.034 0.629

HOOK LAT PAL SPH TIP

TABLE I(B) AVERAGE GRASP MEMBERSHIP FOR SUBJECT POOL (STEADY-STATE) MIMED GRASP CYL CYL 0.753 0.056 0.046 0.043 0.043 0.059 HOOK 0.032 0.804 0.039 0.042 0.028 0.056 LAT 0.021 0.029 0.776 0.092 0.011 0.071 PAL 0.021 0.029 0.082 0.698 0.016 0.155 SPH 0.084 0.039 0.030 0.027 0.741 0.079 TIP 0.035 0.062 0.061 0.179 0.070 0.595

the cylindrical grasp. The same seems to be true of the hook and spherical grasps. It may be the case that intrinsic hand musculature, as opposed to the extrinsic hand muscles analyzed, would show more similarity among these three grasps compared to the other three grasps. The fact that no single grasp has unity self-membership and is therefore not completely distinct seems to suggest that there may be an overlapping control mechanism that the CNS uses to form these hand grasps. This seems to be consistent with work that has suggested that control of hand function is distributed among neurons in the M1 area (i.e. overlapping) [11] instead of concentrated in a specific group of neurons (i.e., non-overlapping). The distribution may be limited though, as all grasp patterns had self-memberships of at least 0.6, both in the transient and steady-state phases. V. CONCLUSION We have used FCM clustering to quantify the electromyographic overlap between six major categories of hand grasps, based on an extrinsic ten-muscle parameter space. These results quantitatively show that no grasp is completely distinct from any other in transient or in steadystate, and that the EMG patterns of the tip and palmar grasps in particular significantly overlap. More subjects will be included to validate these results. Further analysis on amputees or individuals with congenital limb-deficiencies could reveal any potential shift in control paradigm from that seem in intact-limb subjects, and thus prove useful in designing control algorithms for trans-radial prostheses. ACKNOWLEDGMENT The authors would like to thank Dr. Todd Kuiken, Mr. Todd Farrell and Ms. Rebecca Stine for their assistance during the experimental protocol. REFERENCES
[1] A. Keller, C. Taylor, and V. Zahn, Studies to determine the functional requirements for hand and arms prostheses., in Department of Engineering. Los Angeles, CA: University of California at Los Angeles, 1947. H. Gray, Gray's Anatomy, 29 ed. Philadephia, PA: Lea & Febiger, 1973. J. F. Soechting and M. Flanders, Flexibility and repeatability of finger movements during typing: Analysis of multiple degrees of freedom, J Comp Neurosci, vol. 4, pp. 29-46, 1997. N. Bernstein, Dexterity and it's development. Erlbaum, Mahwah, 1996. M. Wiesendanger, Manual dexterity and the making of tools - an introduction from an evolutionary perspective, Exp Brain Res, vol. 128, pp. 1-5, 1999. J. R. Napier, The Prehensile Movements of the Human Hand, J Bone Jnt Surg-Brit Vol, vol. 38, pp. 902-913, 1956. R. S. Johansson and K. J. Cole, Sensory-Motor Coordination During Grasping and Manipulative Actions, Curr Opin Neurobio, vol. 2, pp. 815-823, 1992. N. Kamakura, M. Matsuo, H. Ishii, F. Mitsuboshi, and Y. Miura, Patterns of Static Prehension in Normal Hands, Am J Occ Ther, vol. 34, pp. 437-445, 1980. R. Weir, Design of artificial arms and hands for prosthetic applications, in Standard handbook of biomedical engineering & design, M. Kutz, Ed. Illinois: The McGraw-Hill Companies, Inc., 2003, pp. 32.1-32.61. R. N. Holdefer and L. E. Miller, Primary motor cortical neurons encode functional muscle synergies, Exp Brain Res, vol. 146, pp. 233-243, 2002. A. V. Poliakov and M. H. Schieber, Limited functional grouping of neurons in the motor cortex hand area during individuated finger movements: A cluster analysis, J Neurophys, vol. 82, pp. 3488-3505, 1999. E. Delagi and A. Perotto, Anatomic guide for the electromyographer. Springfield, IL: Charles Thomas Publisher, 1981. J. Bezdek, Pattern recognition with fuzzy objective function algorithms. New York: Plenum, 1981. T. Ross, Fuzzy logic with engineering applications. New York: McGraw-Hill, 1995. L. A. Zadeh, Fuzzy Sets, Information and Control, vol. 8, pp. 338-353, 1965. L. A. Zadeh, Outline of a New Approach to Analysis of Complex Systems and Decision Processes, IEEE Trans Sys, Man, Cyb, vol. SMC3, pp. 28-44, 1973.

GRASP CLUSTER

GRASP CLUSTER

HOOK LAT PAL SPH TIP

Each cell is the average membership across all subjects of the mimed grasp in the respective grasp cluster. The transient portion of the grasp (a) showed the lateral and cylindrical grasps to be the most distinctive (self memberships on the diagonal), while the tip and palmar grasps proved to be the least. Based on the FCM clustering, tip and palmar grasps seem to significantly reciprocate some characteristics electromyographically. This is consistent in the steady-state result. However, in the steady-state result, the hook grasp was the most distinct.

accurate representation of the boundaries between the data, they do not give a sense of any overlap between data points in separate clusters, because each data point is only allowed binary membership in each cluster. FCM clustering is to these other cluster methods what fuzzy logic [15, 16] is to Boolean logic, in that it extends this binary membership to continuous membership. Analysis of the transient phase of the grasp pattern is of interest from a control perspective because it gives insight into the differences the motor system implements to distinguish between grasps. This analysis showed that while all grasps had some level of distinctness, tip and palmar were most similar in terms of muscular activation pattern. The similarity in muscular activity is an expression of the kinematic similarity between these two grasps. In the tip grasp, the index finger opposes the thumb, and in the palmar grasp, the thumb is opposed by both the index and middle fingers. It is of interest that the extent to which the tip and palmar grasps share membership is almost reciprocated during the transient phase. Cursory analysis of the kinetic posture of these grasps suggests that the cylindrical grasp would be most like the spherical or hook grasps, but this does not seem significantly to be the case in either the transient or steady-state phases of

[2] [3] [4] [5] [6] [7] [8] [9]

[10] [11]

[12] [13] [14] [15] [16]

52