During object grasp, a coordinated activation of distal muscles is required to shape the hand in relation to the physical properties of the object. Despite the fundamental importance of the grasping action, little is known of the muscular activation patterns that allow objects of different sizes and shapes to be grasped. In a study of two adult macaque monkeys, we investigated whether we could distinguish between EMG activation patterns associated with grasp of 12 differently shaped objects, chosen to evoke a wide range of grasping postures. Each object was mounted on a horizontal shuttle held by a weak spring (load force 1–2 N). Objects were located in separate sectors of a “carousel,” and inter-trial rotation of the carousel allowed sequential presentation of the objects in pseudorandom order. EMG activity from 10 to 12 digit, hand, and arm muscles was recorded using chronically implanted electrodes. We show that the grasp of different objects was characterized by complex but distinctive patterns of EMG activation. Cluster analysis shows that these object-related EMG patterns were specific and consistent enough to identify the object unequivocally from the EMG recordings alone. EMG-based object identification required a minimum of six EMGs from simultaneously recorded muscles. EMG patterns were consistent across recording sessions in a given monkey but showed some differences between animals. These results identify the specific patterns of activity required to achieve distinct hand postures for grasping, and they open the way to our understanding of how these patterns are generated by the central motor network.
In the action of reaching to grasp, the hand has to be shaped appropriately to match the form, size, and orientation of the object to be grasped. This control of hand preshaping depends on the fine control of hand and finger musculature. At the cortical level, it has been proposed that specific circuits enact a sensorimotor transformation of the object's visual properties into a set of motor commands that will result in the hand shape appropriate for efficient grasp (Jeannerod et al. 1995). This transformation involves a parieto-frontal circuit in which the three-dimensional properties of the object are extracted in area AIP of the posterior parietal cortex (Murata et al. 1996; Sakata et al. 1995, 1997), and the type of movement to be performed is selected in the ventral premotor area, F5 (Jeannerod et al. 1995; Murata et al. 1997; Rizzolatti and Luppino 2001). Relatively little is known about how this visual-to-motor transformation is encoded in motor commands to the hand and digit muscles that shape the grasp. In addition to corticospinal projections from F5 (Galea and Darian-Smith 1994; He et al. 1993), there are dense cortico-cortical projections from F5 to M1 (Godschalk et al. 1984; Matelli et al. 1986; Muakkassa and Strick 1979), and M1 is the major source of descending corticospinal projections influencing hand musculature (Dum and Strick 1991; Maier et al. 2002; Porter and Lemon 1993).
A detailed description of the pattern of muscular activation during grasping is a critical step toward a better understanding of the mechanisms of hand movement control. One of the main issues is the degree of task-specificity of the arm, hand, and digit muscles for the grasp of different objects. This question is of particular importance since it has been reported that human subjects tend to simplify the problem of hand movement control by using a reduced number of postural synergies in grasping a large set of objects of different shapes and sizes (Mason et al. 2001; Santello et al. 1998). In this study, carried out in two rhesus macaque monkeys, we show that there are distinctive and reproducible patterns of EMG activation during grasp of differently shaped objects.
A preliminary report of this work has been published (Brochier et al. 2001).
We trained two purpose-bred adult female macaque monkeys (M. mulatta) to perform grasping movements of differently shaped objects for food rewards. The monkeys will be referred to as M37 (weight, 5.5 kg) and M39 (weight, 5.0 kg). Both monkeys were initially trained to perform the task with the right hand; after 27 wk of recording from M37, she was trained to perform the task with the left hand. The apparatus was in many ways similar to the one previously used by Murata et al. (1996) and consisted of a box divided into lower and upper sections by a half-mirror. The lower part comprised a computer-controlled carousel divided into six separate sectors. In each sector, an object was mounted on a low-friction horizontal shuttle held by a weak spring. A red/green LED was located in the upper part of the box, and its reflection in the mirror was superimposed on the top of the object. The box was positioned in front of the monkey, and the objects were presented one at a time in a pseudorandom order.
The trial sequence was as follows (Fig. 1). The monkey was in darkness during the intertrial interval (usually 1–2.5 s); it was trained to press on two touch or home pads simultaneously with the left and right hands. These were located near the monkey's waist. When both pads were activated by gentle downward pressure from the hands, the object was illuminated, and a red LED was switched on, which was reflected on the object through the half-mirror. After a variable period (1 ± 0.5 s), the LED changed from red to green, and the monkey had to use the trained hand under full vision to reach out, grasp, and pull the object into a position window between 4 and 14 mm from the rest position for 1 s. To displace the spring-loaded object by this amount required a gentle force of 0.9 (4 mm) to 2.4 N (14 mm); the spring behaved in a near linear manner. Hall effect sensors were used to monitor the displacement of the object. A tone was given to indicate that the object was correctly positioned in the window. At the end of this tone, the monkey had to release the object and take a food reward with the other hand. The light was turned off, and the carousel rotated so as to present a new object. Both monkeys were carefully trained to use a specific grasp for each object (Fig. 2), and we used two video cameras to control that these grasp movements were consistent across trials and sessions. The complete training took around 8 mo in M37 and 10 mo in M39.
Each monkey performed a total of 50–100 grasping movements per object during a standard recording session. The 6 objects mounted on the carousel were selected from a total of 12 available objects of different shape and size. The same set of objects was tested in ≥5 different sessions. The dimensions of the different objects (in mm) are given in Fig. 2.
All procedures were carried out in accordance with the UK Home Office regulations. All surgeries were performed under deep general anesthesia, induced with 10 mg/kg ketamine (im), and maintained with 2–2.5% isoflurane in 50:50 O2-N2O. Full aseptic procedures were used. At a first surgery, a custom-designed stainless-steel circular headpiece was securely attached to the monkey's skull (Baker et al. 1999; Lemon 1984). This headpiece was used for head restraint during recording sessions.
Surgical implantation of EMG recording electrodes
In a second surgery, we implanted EMG patch electrodes on ≤12 shoulder, arm, hand, and digit muscles, using the technique described by Miller et al. (1993). This involved suturing a small silastic patch carrying a pair of multistrand stainless steel electrodes onto the surface of the muscle. The entire implant was prepared presurgically and electrodes were connected subcutaneously to a 25-way miniature d-connector supported in a dacron patch and implanted in the skin over the back of the monkey. The muscles implanted were: first dorsal interosseous (1DI), abductor pollicis longus (AbPL), thenar (Th), abductor digiti minimi (AbDM), palmaris longus (PL)39, flexor digitorum sublimis (FDS), flexor digitorum profundis (FDP), extensor digitorum communis (EDC), extensor digitorum 4,5 (ED45), flexor carpi ulnaris (FCU), extensor carpi ulnaris (ECU)39, brachioradialis (BR)37L, anterior deltoid (AD). In M37, after recordings from the right side were completed, a second EMG implantation was carried out for the left side. Muscles implanted in only one monkey are indicated with the superscripts 37L and 39 for monkeys 37 (left arm) and 39, respectively.
After each surgery the monkeys received a full course of antibiotics (oxytetracycline, 20 mg/kg im; Terramycin/LA, Pfizer) and analgesic (buprenorphine, 10 μg/kg im; Vetergesic, Reckitt and Colman).
EMG signals were amplified 2,000 times, high-pass filtered at 30 Hz (Neurolog EMG amplifier, NL824, Digitimer), and sampled at 5 kHz using a A2D interface (PCI-6071E, National Instruments). These signals were recorded together with key behavioral events, including the timing of home pad release (HPR), first movement of object, and an analog record of object displacement.
Cross-talk between EMG channels
A certain level of physical cross-talk between pairs of EMG recordings can result from volume conduction in the surrounding tissues. Such cross-talk would obviously compromise the independence of the EMG recordings that were made during the task. We investigated the level of any cross-talk present by calculating the cross-correlation between every pair of simultaneously recorded unrectified EMGs over a period of 100 s. For this analysis, a peak value of r of 1 at zero-lag would indicate identical, completely redundant EMG recordings, while an r value close to 0 would correspond to uncorrelated, independent signals. Cross-talk was negligible for the majority of the muscle pairs; in 134 of 147 pairs analyzed, r was <0.1. In a few cases (12/147), it peaked below 0.2. It reached a value of 0.3 in only one pair of recordings (FDS*FCU in monkey M37L), where the electrodes were physically very close together. In general, the low correlation values confirm that the EMG signals recorded from different muscles were independent.
Confirmation of EMG recording sites
In both monkeys we were able to confirm the precise location of each pair of recording electrodes over the selected muscle in the left arm/hand by electrically stimulating through the back connector. Monkeys were deeply anesthetized during this procedure. This evoked twitches of the implanted muscles with low stimulation voltages (∼0.5 V). At the end of all experimental procedures, M37 was given an overdose of anesthetic (Lethobarb, 100 mg/kg ip) and perfused transcardially. Dissection of both left and right arms was carried out to further confirm the electrode location. M39 is still alive.
Time-course of EMG activation during the task
The results are based on recordings from five sessions in M37 using the right arm (M37R) and five in M37 using the left arm (M37L). Ten sessions were recorded in M39 using the right arm, five with one group of objects (M39Rg1) and five with a second group (M39Rg2). Off-line, EMGs were rectified, and averaged activity was calculated for performance on each object across all the trials within that session, using the onset of HPR as a reference event. Typical averages from a recording session with M39 are shown in Fig. 3 for three muscles (Th, FDS, EDC) and six different objects (group 2 of M39). The black vertical line on the left marks the time of HPR and thus the onset of the reach to grasp movement. The black vertical line on the right corresponds to the onset of object displacement and its 95% confidence interval (2 gray vertical lines). These intervals suggest that the duration of the movement was object-specific and showed little variability across trials. All three muscles were strongly activated during the movement but with distinct task-related patterns. The thenar muscles (Th) were mostly active during grasp and just before object displacement. In contrast, FDS and EDC activity peaked earlier during the grasp; EDC was the only muscle that showed an initial burst of activity just before HPR.
For a given muscle, both the pattern of activation and the depth of modulation varied with grasp of different objects. For instance, thenar muscles were strongly activated for both the precision grip and the large disc but with an earlier onset of activity for the precision grip; these muscles were almost inactive during grasp of the ring. The tuning of EMG activity to object grasp was much sharper for EDC, with a pronounced activation of this muscle for the precision grip, but rather similar patterns for some of the other objects.
To analyze these between-object differences, we computed a normalized time scale to compare the muscle activity at specific epochs during the movement. This was done by measuring the time difference between HPR and the onset of object displacement for each trial (usually between 300 and 500 ms depending on the object grasped) and by dividing this period into 10 consecutive epochs. The EMG activity was averaged within each epoch and across trials for each object. The averaged activities of thenar and EDC for the objects showing the strongest and weakest averaged level of activity across epochs E1 to E10 are shown in Fig. 4, along with the between-object variance computed across all six objects. The sharp increase in activity in the third epoch after HPR was combined with a parallel increase of between-object variance, suggesting an early differentiation of muscle activity during the reach-to-grasp movement. Interestingly, the earliest burst of activity of EDC around time 0 was not reflected in the variance measure and is probably due to the fact that the extension of the fingers preceding HPR was common to all trials.
Figure 5 presents the time course of EMG activation in each of the 12 muscles, superimposed for the six different objects. The plots in Fig. 5 confirm the impression given by the raw data shown in Fig. 3 that the time course of muscle activation differed markedly from one muscle to another. For instance AbPL and ED45 were more active midway during the reaching movement, whereas FDS and FDP were more active toward the onset of object displacement. The contrast was particularly clear-cut for AD and 1DI, which were activated around HPR and the displacement onset, respectively.
The EMG activity in different muscles also showed marked variation in the degree of object-related tuning. EDC was highly tuned for the precision grip while PL was tuned for the small ring and FCU for the large cone. For the other muscles, the activity was tuned for two or three objects (AbDM or ECU) or more widely distributed across the six objects (1DI, Thenar, ED45). Only the proximal muscle anterior deltoid (AD) showed little or no sign of object differentiation. Some muscles showed peak activity associated with different objects at different times during the reach-to-grasp action. For instance, the thenar muscles were more active for the precision grip in the early part of the movement but showed more pronounced activity for the disc and the cone just prior to displacement onset.
Figure 6 analyses in more detail the degree of object-specific tuning of EMG activity for each recording in the two epochs shown in Fig. 4: halfway through the movement (epoch E5: 40–50% of movement time) and just before object displacement (epoch E10: 90–100% of movement time). We also quantified the object-tuning by calculating a “selectivity index” for each of these two epochs, as defined by Rainer et al. (1998) where n is the number of objects, ri the activity for object i, and rmax the maximum ri. S is close to 1 when the muscle is active for only one object and close to 0 when the muscle shows the same level of activity for all six objects.
As suggested in Fig. 5, the object-specificity varied across the 12 muscles and between the two selected epochs. In epoch E5, the specific tuning of EDC and PL, for the precision grip and the small ring, respectively, is expressed by the high values of the selectivity index for these muscles (0.8 and 0.67 for EDC and PL, respectively). On the other hand, the less differentiated activity of AD was associated with an index of only 0.27. For all the muscles, the object tuning in E10 changed in comparison with E5: it was not possible to predict from the EMG activity in E5 how the muscles will be tuned in the later part of the grasp. These clear changes in EMG tuning during the task were assessed by calculating for each object the coefficient of correlation between the level of activity of all the muscles in E5 and E10. The correlation of EMG activity between these two epochs was weak and reached significant levels (P < 0.05) in only a small number of cases (objects 5 and 3 in M39Rg1 and M37L, respectively, and objects 2, 3, and 5 in M37R).
These observations disclose the complexity of muscle coordination underlying the control of grasp movements. We investigated whether or not these complex patterns of EMG activity were sufficiently consistent within and between recording sessions to allow differentiation between object-specific grasp on the basis of the EMG recordings.
Object-specific activation of muscles during grasp
For this part of the analysis, the EMG activity was averaged over blocks of 10 trials randomly sampled among all the trials for a given object. Within a given recording session, 5 of these 10-trial blocks were averaged for each object. We normalized the variations of EMG activity by dividing the EMG activity in every epoch by the peak of EMG activity measured among all blocks, epochs, and objects. This normalization process used the same amplitude scale (0–1) and allowed comparison between levels of EMG activity across muscles.
To discriminate between the different objects on the basis of the EMG recordings, each object was represented as an n-dimensional muscle vector (NDMV), with each dimension describing the averaged and normalized level of activity for a given muscle in a block of 10 trials. These NDMVs were computed epoch-by-epoch to analyze in detail the differentiation of object-related EMG patterns with time.
In Fig. 7, two of these NDMVs are represented on polar plots for epochs E5 (continuous line) and E10 (dotted line) in one of the recording sessions from M39Rg2. These plots indicate that the overall level of EMG activity varied from object to object, with the plate and the ring requiring less activation than the four other objects. In epoch E5, all the patterns were characterized by an elongated shape pointing toward the AD, indicating the strong activation of this muscle in this early part of the task. However, for other muscles, it is clear that the activity in E5 was already differentiated across objects. PL was strongly activated for the small ring, EDC for the precision grip, and AbDM for the large cone. Although the NDMVs for the three objects on the top row showed some similarities, they exhibited some important differences. For instance, the activation of AbDM was much stronger for the large disc and the activation of ED45 was more specific to the small cylinder.
The EMG patterns in epoch E10 were highly differentiated from those in epoch E5, and they also varied from object to object. The NDMV for the small cylinder and the large cone, for which the grasps adopted were in some ways similar, shared some features: there was strong activation of 1DI and FCU with absence of activation of EDC. However, for these two objects, the levels of EMG activity were strikingly different for other muscles. There was stronger activation of FDP for the cylinder but of Th and ED45 for the cone. Thus rather superficial similarities in grasp posture were not good predictors of EMG pattern. Rather, it appears that the selected objects evoked a specific combination of activity in the different muscles, such that the NDMV patterns were distinct for all the objects.
A cluster analysis, as previously used by Poliakov and Schieber (1999) was applied to assess the degree of similarity between these NDMVs. This was firstly done within each recording session by comparing the NDMVs computed across blocks of 10 trials for the same object. The Euclidian distance (D) between a pair of NDMVs Vi and Vj was computed as follows Once calculated, we linked together into clusters the pairs of trials showing the shortest distance between them. These newly formed clusters were in turn clustered and so on until all the trials in the session were linked together in a hierarchical tree. Therefore in this tree, each NDMV is close to the other NDMVs with similar activity patterns.
To quantify the efficacy of this method for grouping together the NDMVs related to the same object, a percentage of correctly sorted objects (Cs) was calculated on the basis of the order of the NDMVs in the hierarchical tree. Cs was defined as follows where Gract is the actual number of groups in which one to five NDMVs made from a given object were next to each other in the hierarchical tree; Grmax is the maximum number of possible object-related groups when none of the NDMVs were next to a NDMV related to the same object (i.e., Grmax was equal to the total number of NDMVs); and Grmin was the minimum number of possible groups when all the NDMVs from each object were correctly grouped together. Cs was 100 if all the objects were correctly sorted, and 0 if none of the NDMVs from same object were grouped together.
Figure 8A shows the result of this analysis when applied to one session of M39Rg2 in epoch E10 just before the onset of the object displacement. The matrix displays the D between every possible pair of NDMVs with blue and red, representing the NDMVs that had the highest and lowest levels of similarity, respectively. In this session, and for this selected epoch, the distance was the shortest between all the NDMVs corresponding to the same object. Therefore these NDMVs had been grouped together by the clustering process, and the resulting Cs was 100%. In addition, the color code indicates that the degree of similarity varied between one pair of objects and another; the distance was shorter within than between the pairs of objects 3–4, 1–5, and 2–6.
Consistency of object-related muscle patterns across recording sessions
The cluster analysis was used to assess the reproducibility of EMG patterns across five different sessions with the same set of objects. For this part of the analysis, the EMG patterns were averaged for each object across all the trials within a single session, and these averaged values were used to compute the NDMVs. The 30 NDMVs thus created (5 sessions × 6 objects) were compared following the same principles as for within session comparisons.
The color matrix in Fig. 8B shows that, in epoch E10, the NDMVs corresponding to the same object are consistently grouped together. Again, larger groups are formed by object pairs 3–4, 1–5, and 2–6. In epoch E5, the NDMVs were also highly consistent between sessions, but they were organized differently in the similarity matrix (data not shown). Objects 1–3 were close to each other but were together distinct from the other three objects. Between-session consistency was also observed for the other set of objects (M37L, M37R, and M39g2). Overall, these results suggest that the EMG patterns used by the monkeys to grasp different objects were very consistent over time.
Time course of between object discrimination
The results described above suggest that the EMG patterns were object-specific at particular time points (i.e., in epochs E5 and E10). A critical issue concerns how much variation there was in this object specificity as time elapsed through the reach-to-grasp action. This question was addressed by running the cluster analysis epoch-by-epoch from the initiation of the movement (HPR; Fig. 9) to the object displacement onset (DO). The results of this analysis are presented in Fig. 9.
In this diagram, the abscissa indicates the 10 epochs between HPR and DO plus 5 epochs before and after these two events defined in the same normalized time scale. The ordinate shows the proportion of correctly sorted objects (Cs). Each plot represents the data set computed for one monkey and the set of objects it grasped (see Fig. 2). The points plotted on the line indicate the Cs in the given epoch calculated first within each independent session and averaged across five sessions with the same set of objects. The plots show a sharp rise in the Cs as early as in the third epoch after HPR. This observation can be related to the sharp increase of between-object variance observed for some of the muscles (see Fig. 4). By epoch E5, the value of Cs is close to its maximum for the four sets of objects.
It is noticeable that despite the fact that the EMG patterns were significantly different in epochs 5 and 10 (see Fig. 7), they remained strongly object-specific throughout the intermediate epochs (E6–E9). The object-specificity is also maintained in the latest part of the grasp following the displacement onset. It is also clear that monkey M37 tended to anticipate the object to be grasped well before releasing of the home-pad switch, as indicated by the Cs values well above chance level for M37L and M37R. Indeed, this monkey occasionally tended to close the home-pads using downward pressure of the forearm rather than by pressing with the hand itself. In these situations, the hand and fingers were free to move before the actual release of the home-pad and could anticipate the posture. This problem was not observed in monkey M39 who consistently used her hands to depress the home-pad switches.
Muscle dimensionality for object discrimination
The histograms in Fig. 6 suggested that the dimensionality of the muscle space may be less than the actual number of recorded muscles. For instance, the activity of AbDM and ECU seems to closely co-vary in both epochs E5 and E10. This redundancy might limit the amount of information available for the clustering analysis. In this respect, one can question whether a subset of variables describing most of the variance in our EMG sample would be sufficient to differentiate between the object-related EMG patterns. A principal component analysis combined with the cluster analysis was used to answer this question. First, we constructed a matrix, X, comprising 600 × 12 values, to analyze the co-variance between pairs of muscles. In this matrix, 600 represented 20 epochs × 5 averaged blocks of 10 trials per object × 6 objects, and 12 represented the number of muscles from which EMG was recorded.
The 20 selected epochs were the 10 epochs between HPR and OD onset, plus 5 epochs before HPR (i.e., to the left of HPR in Fig. 9) and after DO (i.e., to the right of DO in Fig. 9) computed on the same normalized time scale. The 12 principal components (PCs) were extracted from the eigenvectors and eigenvalues of the covariance matrix of X. The principal component analysis showed that the first three PCs accounted for about 81% of the variance in our data. Figure 10A plots the composition of the first two PCs against each other for monkey M39Rg2. As expected from Fig. 6, AbDM and ECU were very close to each other in this two-dimensional space, as were the muscle pairs EDC-PL, FCU-FDS, AbPL-ED45, and 1DI-FDP. The more proximal AD is very distant from all the other muscles.
We then tested whether these principal components would allow us to distinguish between the different objects using a reduced number of variables to describe the animal performance. This was done by computing, for each epoch, a multidimensional vector in the coordinate system defined by the principal components. A cluster analysis as described above was applied to compared these newly formed PCs vectors. The dimension of these vectors was defined arbitrarily by selecting 1, 3, 5, or the 12 principal components for the analysis. The order of the principal components selected to construct the vectors was either descending (starting from the 1st PC with the highest eigenvalue to the 12th PC) or the reverse, ascending order.
Figure 10B shows that using the first PC only, just over 60% of the 10-trial averages could be correctly identified as corresponding to the same object. Cs was close to its maximum value when a five dimensional PCs vector was used for the discrimination; the remaining seven PCs accounted for <2% of the total discrimination. Interestingly, however, when the last PCs were used first for the discrimination (Fig. 10C), five of them were enough to account for a Cs of nearly 90% in epoch E5. This result suggests that even if single PCs account for a very small fraction of the variance in our data, they are sufficient to define finely tuned patterns of activity for each object when coordinated with other PCs.
Between-monkey variability in object-related muscle patterns
During training, special care was taken to ensure that both M37 and M39 were using a similar type of grasp for each of the object. The overall similarity of the object-related grasp movements for the two monkeys was confirmed by comparing the video-images of their hand postures during grasp (cf. Fig. 2 for M39). We investigated whether these similarities in the hand posture would be reflected in the object-related muscle patterns. This comparison was limited to epoch E10 in which the monkeys are about to pull the object. Figure 11 shows a superimposition of the E10 NDMVs of M37R (continuous line) and M39Rg2 (dotted line) for the first three objects shown in Fig. 2. These plots suggest that the global shapes of the NDMVs shared some similarities for the small cylinder but do not superimpose as precisely for the other two objects (large disc and small plate). For instance, the activation of the extensors (EDC, ED45) was stronger in M39 than in M37 when grasping the disc. In contrast, M37 activated the intrinsic hand muscles (mainly 1DI and AbDM) to a greater extent during grasp of the small plate than did M39.
We used the clustering method to analyze whether, despite these differences, the E10 NDMVs for a given object would be grouped together when originating from the two different animals. This analysis showed that, for all the objects but the small ring, the NDMVs from M37R and M39Rg2 were identified as corresponding to distinct and independent grasps.
In this study, a cluster analysis was used to show that, in overtrained monkeys, there are distinct patterns of hand and arm muscle EMGs related to grasp of different objects. These patterns provide an accurate estimate of the degree of dissimilarity between the different grasp movements monkeys used to perform the task. These data also provide useful information on the consistency of these EMG patterns across monkeys and recording sessions.
The design of the experiment was such as to require the monkey to perform a rich variety of grasp movements to grip and displace the different objects. Our approach was to try to understand how this repertoire of different grasps is reflected in the activity of a representative sample of arm, hand, and finger muscles. To make a systematic analysis, we were particularly careful to train the monkeys always to grasp a given object in the same way. The total envelope of EMG activity recorded was, of course, related to different phases of the task, including reaching, hand shaping, grasping, and pulling of the object. The load force resisting object displacement (spring constant) and the frictional properties of the object surface were the same for all objects. These are two of the major factors determining grip force in humans and monkeys (Picard and Smith 1992; Westling and Johansson 1984), and we have assumed that the grip force exerted by the monkey was similar for each object. Actually monitoring the grip force would be technically difficult, because it would have been distributed very differently for different objects and would have involved different digits and opposition spaces (Jenmalm and Johansson 1997). Differences in grip force and its distribution across objects will of course be reflected in EMG activity and form part of the object-based specification of the command to hand and finger muscles during grasp (Maier and Hepp-Reymond 1995a).
Comparison of object-related EMG patterns
The polar plots in Fig. 7 show the distinctive pattern of EMG activation for each of the six objects tested at two different points in time within a given session. We consider first the relationship between these EMG patterns and the hand postures adopted by the monkey to grasp the objects. Our analysis shows that the pattern of muscle activation was clearly different for each object, both while the monkey was shaping the hand during reach and while generating the force required for the pull (epochs E5–E10, Fig. 9). The level of activity in a given muscle could differ sharply during adoption of rather similar hand postures, as shown for the small cylinder and the large cone in Fig. 7. This may reflect, for example, differences in grasp aperture between index finger and thumb. A large and representative sample of muscles was required to get a fuller perspective of grasp-related EMG activity.
We conclude from our cluster analysis that there were critical differences in the grasps used that could not be accounted for by a simple classification of the postures adopted. A number of different factors may explain discrepancies between hand postures and EMG patterns. First, it is known that during finger movements, an important part of the muscle activation is not involved in producing motion but in compensating for the interaction torques between the different finger segments (Darling and Cole 1990). Even minor differences in the posture of the hand at object contact could alter these interaction torques and result in significant changes in the task-related muscle activity. Second, significant changes in EMG activity are evoked by the actual contact between the fingers and the object surface (Collins et al. 1999). Given the contrast in the distribution of digit surface contact across objects, it is likely that such contact effects contributed to the differences observed in grasp-related EMG pattern.
Overall, our results indicate that, for reach-to-grasp movements, the detailed EMG pattern provides an accurate characterization of the motor output requirements over and above that given by hand posture around the object. There is no doubt that such EMG data provide essential information to interpret the activity of cortical neurons when recorded in a similar type of task (Murata et al. 1996, 2000; Sakata et al. 1995). For instance, Murata et al. (2000) were puzzled to find that the population of grasp-related neurons in the parietal area AIP was activated in a closely related manner for grasp of two objects with different hand postures (the plate and the ring). Such a result makes more sense in light of our data showing that, at the muscle level, the distance between these two objects is very short (Fig. 8B).
Multidimensional coding of the EMG patterns
Our results show that each of the objects can be accurately discriminated if the whole set of recorded muscles is included in the identification process (Fig. 9). The principal component analysis indicates a certain level of co-variation for some muscle pairs and suggests that our analysis could be simplified by reducing the number of variables required to distinguish between the different objects. However, we show that a three-dimensional vector extracted from the first three PCs is not sufficient to adequately characterize each single object (Fig. 10B), confirming that a multi-dimensional control is the basic unit that is used here (Buys et al. 1986; Holdefer and Miller 2002; Porter and Lemon 1993; Schieber 1995). It is of interest to note that a combination of PCs with the lowest eigenvalues could also be used to discriminate between the different objects (Fig. 10B). This observation emphasizes the fact that all 12 dimensions contribute to the definition of the global patterns related to each objects. The sample of muscles used in our study is still but a small proportion of the total number of muscles involved in the task and the dimensionality of the EMG patterns underlying grasp might be even greater than reported here when the whole hand and forearm musculature is considered.
Schieber (1995) has previously shown that independent finger movements are produced by the combined activity of several muscles. It is also known that both single cortico-motoneuronal cells and populations of such cells influence small but specific groups of synergistic finger muscles that help to form the “building blocks” or “primitives” of different grasps (Buys et al. 1986; Jackson et al. 2003). These mechanisms might be active in our experiment in which the target objects were designed to elicit a set of distinct finger movements and grasping postures.
In this study, we have not attempted to address the issue of the organizational principles underlying grasp-related patterns of EMG activity. However, some critical properties of these EMG patterns can be considered from the activity histograms in Fig. 6. First, there were clear differences in the timing of activity across the sampled muscles. In general, the intrinsic hand muscles (1DI, Th, and AbDM) and the long flexors (FDS and FDP) were more active in the later part of the task when the digits were already in contact with the object surface (gray histograms on Fig. 6). This would confirm the hypothesis that these muscles are mainly involved in the production of isometric forces (Maier and Hepp-Reymond 1995a). The EMG activity of the other tested muscles was more evenly distributed across the time course of the task. For all the muscles, the lack of correlation between the patterns of activity in epochs E5 and E10 suggest that they are differently involved in the control of hand posture during reaching and in the adjustment of the grip forces during pulling. Only the AD was exclusively active during the early reaching part of the movement.
Second, some coherent spatio-temporal relations in muscle activation can be seen in Fig. 5. EMG activity in AbDM, AbPL, ED45, and ECU, and to a lesser extent for 1DI and FDP, showed a certain level of co-variation during the task. These relations were more clearly identified by the principal component analysis in which these groups of muscles are closely related along the first and second PCs. Such synergistic patterns of EMG activity have been described previously in the control of precision grip (Maier and Hepp-Reymond 1995b) and are likely to reveal a more general principle in the control of grasping movements. It has been hypothesized that the activity of primary motor cortical neurons encodes the activity of functionally relevant groups of muscle (Bennett and Lemon 1996; Buys et al. 1986; Holdefer and Miller 2002).
Consistency of recorded EMG activity within and between individual monkeys
The cluster analysis suggested that the EMG pattern related to a given object was reproducible across sessions. However, comparison of the NDMVs for the same object in the similarity matrix (Fig. 8, A and B) shows some variability in these patterns within and between sessions. This variability was most marked in the comparison of the NDMVs from the two monkeys (Fig. 11).
There are several possible sources of variability in the EMG pattern. We trained the monkeys to use the same consistent grasp posture for a given object, and we used regular video monitoring to control the monkeys' performance. Nevertheless, we cannot exclude some variability in the actual movement produced and posture adopted to perform the task. As already mentioned, such changes in hand postures also affect the intersegmental interaction torques and are likely to result in significant variations in EMG activity (Darling and Cole 1990). In addition, since the number of muscles involved in performing the task exceeded the number of tested objects, one can argue that the combination of muscle activity used to produce a given hand posture could differ slightly across trials and sessions within the same monkey (Loeb 1993; Schieber 1995). This might also explain the differences in EMG patterns observed between the two monkeys (Fig. 11), although it is also possible that it was caused by variation in limb morphometry or by electrode placement in the two animals, particularly in the large multidigit muscles (such as FDP) in which distinct functional subdivisions have been identified (Cheng and Scott 2000; Schieber 1993).
This study suggests that the patterns of EMG activation in a representative sample of arm and finger muscles provide a reliable representation of the monkey's motor behavior during grasping of differently shaped objects. This type of information is of critical importance for the interpretation of central mechanisms controlling grasp.
This study was supported by the Wellcome Trust, Medical Research Council, and European Union project QLRT-1999-00448; Cortico-spinal Modelling.
We acknowledge the collaboration of Professors Giacomo Rizzolatti and Vittorio Gallese (Institute of Human Physiology, University of Parma, Italy). Many thanks to Dr. A. Jackson for help and to Professor M. Maier for comments on the manuscript. Technical assistance was provided by H. Lewis, S. Shepherd, V. Baller, E. Bye, and J. Turton.
The costs of publication of this article were defrayed in part by the payment of page charges. The article must therefore be hereby marked “advertisement” in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
- Copyright © 2004 by the American Physiological Society