I. INTRODUCTION

Top Previous Next References

The term proprioception was coined by Sherrington (299) to describe the sensory information contributing to a sense of self position and movement. The relationship of proprioception to unconscious or more automatic functions has traditionally distinguished it from kinesthesis, or the conscious sense of position and movement. Proprioception has also been associated with a distinct class of sensory receptors, most notably those found in the muscles and related deep tissues, while kinesthesis has been more closely associated with joint and cutaneous receptors (45, 78, 127). The framework we propose here considers how proprioceptive sensory information is organized as it relates to a sense of body position and movement, but without any judgement about which sensory receptors may or may not be involved.

Current views of proprioception at the spinal level have, however, been strongly influenced by our understanding about the sensory receptors that contribute to it (for reviews, see Refs. 146, 199, 290). Some of these, the muscle spindles for example, have been studied extensively so the information they transmit to the nervous system is known in great detail (for example, see Refs. 77, 137, 138, 196, 218, 219, 272, 277-280, 296, 325). This knowledge, combined with our knowledge of receptor terminations and functional connectivity, has led to a general understanding of spinal proprioception as a central expression of the sensory receptor input which in turn details local parameters like muscle length and muscle force. The intent of this review is to examine this perspective in an historical context, and through that examination to provide a different framework for understanding spinal proprioception that may be more in tune with current views of sensorimotor processing in other central nervous system (CNS) structures.

	II. IMPORTANCE OF PROPRIOCEPTIVE FEEDBACK FOR MOTOR ACTIONS

Top Previous Next References

Contemporary thinking about motor control has it that many actions belonging to a daily repertoire of motor activity may actually be controlled in a feed-forward manner, i.e., without a direct contribution from sensory feedback (85, 159, 162, 176, 209, 229, 333). Thus a loss of proprioceptive feedback does not prevent movement or the ability to carry out motor tasks. Polit and Bizzi (268, 269) demonstrated this by showing that deafferented monkeys could perform a reach-to-target task successfully without visual aid. However, the behavior was not normal, suggesting further that sensory information may play an important regulatory role even for such over-learned movements as these.

This issue has now been investigated in a number of different behavioral contexts and with diverse experimental approaches. Arm movements, particularly reaching movements, have continued to provide a preferred experimental paradigm for investigating the contributions of sensory information in guiding voluntary motor actions (106, 167, 168). For example, studies by the Ghez laboratory on deafferented patients lacking proprioceptive feedback have provided a number of relevant observations. These patients made large directional and amplitude errors when performing a reaching task without visual feedback. The errors were associated with differences in velocity and acceleration for movements in different directions and, according to the authors' interpretation, might have resulted from a failure to take the inertial properties of the limb into account in programming the initial trajectory (113). Because direction and amplitude errors were partially compensated by allowing patients to view their arm before or during the movement, the authors proposed that visual feedback could assist proprioception or even substitute for it in some way by updating an internal model of the arm. (105).

A variant on the traditional reaching task was used in another study with a similar group of patients to show that a lack of proprioceptive information might also be responsible for deficits in interjoint coordination (107, 287). More recently, Sainburg et al. (286) speculated further that "control of intersegmental dynamics is normally dependent on proprioceptive information to update and maintain neural representations of the musculoskeletal system." They found that sensory feedback was also required for adaptation to novel intersegmental dynamics. Thus, in agreement with other studies of this genre (64, 110, 183, 297, 317), they concluded "the nervous system uses sensory information to develop and recalibrate internal models of the musculoskeletal system itself."

These studies, among an increasingly large number on this subject, clearly outline the importance of a proprioception both as a continuous feedback process and also as a component in the central processing of voluntary movements (see Refs. 126, 161, 168, 222, 223, 262, 273 for reviews). Although many of the higher order sensory functions contributing to internal models or intersegmental coordination have been attributed to supraspinal structures like the cerebellum (22, 109, 175, 294, 319, 320), we would like to consider here that an organizational framework for proprioception found at spinal cord levels may also be consistent with such functions.

	III. SPINAL PROPRIOCEPTION: HISTORICAL OVERVIEW

Top Previous Next References

Much of the early work on spinal mechanisms begun by Sherrington and his colleagues and contemporaries (62, 192, 204) was concerned with global aspects of reflex behavior and how sensory and motor mechanisms contributed to them. The emphasis was on behavioral relevance, and it developed from Sherrington's concept that reflexes formed a basis for a hierarchical organization of the sensorimotor system. This top-down investigative approach was largely superceded however following the introduction of the microelectrode in the 1950s. The technological breakthrough provided by the microelectrode allowed investigators to examine functional anatomy in great detail at the single-cell level, and it thereby focused research efforts toward the behavior of individual neurons and the specific synaptic interactions between them.

Two basic concepts from sensory physiology also contributed significantly to the formulation of a general view about the central organization of proprioception. The concepts were the topographical projections of sensory information to and within the CNS, and receptor or modality specificity within the CNS (235, 236, 265, 314). As a consequence of the importance placed on these concepts, much of the effort to understand spinal proprioception was directed at the receptors themselves and at elucidating their specific connections in the spinal cord. The result has been that relatively little attention was paid to understanding whatever composite information might be represented by ensembles of diverse receptors. Moreover, this receptor-specific approach persisted even though most of the synaptic targets of the sensory receptors were found to receive convergent inputs from receptors having different locations and serving different modalities (1, 2, 23, 52, 60, 65, 71, 89, 148, 182, 234, 247, 256, 289). Some better known examples are the convergence of different types of cutaneous receptors onto "wide dynamic range" cells of the spinothalamic system (226) and the convergence of muscle spindle and tendon organ afferents onto common sets of interneurons in specific reflex pathways (224).

Nevertheless, the concept of topological projections has provided a useful framework for understanding function in a number of sensory systems. In cases where the sensory receptors exist in a two-dimensional array, such as the retina or the skin, central organizations can be readily analyzed in terms of a topological mapping that conserves nearest neighbor relationships (e.g., Refs. 43, 44). It has been tempting, therefore, to extend this concept to proprioception, even though its receptors may not be so simply arrayed. Muscle receptors for example are distributed throughout the three-dimensional bulk of individual muscles, and their distribution is not uniform across muscles. Nevertheless, some simple reflex organizations suggested that an appropriate "topology" for proprioception might be defined by the organization of muscles about a single joint, i.e., the agonists and antagonists. For example, it is well established that spinal motoneurons receive direct synaptic input from muscle spindle receptors in the muscle innervated by the motoneuron as well as from synergistic agonists (for reviews, see Refs. 47, 93, 128, 218). Moreover, a consistent reciprocal relationship between receptors and their projections to antagonist muscles is also well established. (141, 193; for reviews, see Refs. 19, 139).

This precise muscle-by-muscle organization may not, however, represent a general organizational pattern. For example, even simple reflex behavior involving more than one joint shows clear interactions across joints (189). Some of these interactions may result from specific reflex pathways such as those linking muscles acting at different joints (49, 239, 242, 244, 295), or the excitation of extensor motoneurons throughout the hindlimb by ankle extensor afferents (125). Heteronymous interactions may also involve presynaptic mechanisms as suggested by studies of afferent-induced modulation of the soleus H reflex during passive cycling movements (230, 231) and stretch of the quadriceps muscle (232).

The idea that the selective sensitivity or specificity of sensory receptors carries over to the CNS has also influenced how experimental results have been interpreted. This principle is emphasized by the concept of sensory modality, which may be thought of as a class of sensations sharing a common stimulus quality. Although such qualities may not precisely correspond to the specificity of sensory receptors, it is often viewed in those terms. For example, until relatively recently, it was considered well established that position sense and kinesthesia are attributable to joint receptors rather than to muscle receptors (e.g., Ref. 236). Joint-angle sensation was believed to result from the central processing of information from joint receptors that specifically encode joint angle (48, 82, 241, 331). However, it is now clear that muscle spindles, once thought to encode exclusively individual muscle lengths, are also major contributors to the kinesthetic sense of position and movement (46, 51, 219-221, 326, 327). This example illustrates how an ensemble of information from diverse sensory receptors may contribute to a given "modality."

	IV. A MODEL TO STUDY SPINAL PROPRIOCEPTION: THE DORSAL SPINOCEREBELLAR TRACT

Top Previous Next References

Two major groups of spinocerebellar neurons contribute to dorsal and ventral spinocerebellar tracts (DSCT and VSCT, respectively), and together they provide the major direct sensory projection from the hindlimbs and lower part of the trunk to the cerebellum. Sensory information from the hindlimbs is also relayed by indirect spinoreticulocerebellar pathways and through at least two olivocerebellar pathways, the direct spinoolivocerebellar and the indirect spinoreticuloolivocerebellar. We consider here only the dorsal spinocerebellar neurons, since they are most likely to represent the purely sensory aspects of the proprioceptive information directed centrally. The ventral spinocerebellar neurons have been shown to encode some form of motor or premotor signal derived from descending pathways in addition to proprioception (13-15, 198, 202). A parallel system serving the forelimbs includes the direct cuneocerebellar and rostral spinocerebellar tracts (118, 258) and other indirect pathways via the lateral reticular nucleus and the inferior olive.

The cells of origin of the DSCT are classically described as residing in Clarke's column in Rexed's lamina VII of the lumbar and thoracic spinal cord segments (Fig. 1A; Refs. 30, 195, 255, 314, see also Ref. 257 for a review). However, the development of anatomical tracers like wheat germ agglutinin (WGA)-horseradish peroxidase (HRP) led to the identification of several other groups of neurons that also belong to the DSCT. These cells are located throughout the intermediate and dorsal laminae of the thoracic and lumbosacral segments of the spinal cord (72, 210, 214, 217). DSCT axons terminate as mossy fibers in the cerebellar cortex of lobules I-V in the anterior lobe and in the posterior lobe vermis and paramedian lobe (117, 213, 215, 216). In addition to the cortical projection, there is recent evidence that DSCT fibers also terminate in the medial and interpositus cerebellar nuclei (211, 212).

View larger version (31K): [in this window] [in a new window]   Fig. 1. Neurons of the dorsal spinocerebellar tract (DSCT). A: camera lucida reconstruction of a Clarke's column DSCT neuron filled intracellularly with horseradish peroxidase (HRP). These are large multipolar neurons with a dendritic arborization that extends beyond the boundaries of Clarke's column (dashed line). [From Walmsley and Nicol (324).] B: muscle receptor axons form giant synaptic terminals on Clarke's column neurons. The morphology of a giant synapse is illustrated by this reconstruction.[Modified from Szentagothai and Albert (315).] C: the giant synapses give rise to large amplitude unitary excitatory postsynaptic potentials (EPSPs). An example is illustrated in this intracellular recording from a DSCT neuron, showing a train of 2- to 4-mV unitary EPSPs evoked by a 50-g load placed on the soleus tendon. [From Eide et al. (74).] D: maximal stimulation of the low-threshold afferents in a muscle nerve may activate from 10 to 20 giant synapses synchronously. The record illustrated here shows a 20-mV maximal EPSP evoked in the same neuron as C by stimulation of the gastrocnemius-soleus nerve. [From Eide et al. (76).]

	V. EARLY STUDIES OF DORSAL SPINOCEREBELLAR TRACT

Top Previous Next References

The earliest results showed that the synapses between muscle receptors and DSCT neurons, unlike those with motoneurons, were "giant synapses" (see Fig. 1B, Ref. 315) that could faithfully follow high-frequency repetitive stimulation of low-threshold muscle sensory fibers (Ia afferents; Refs. 132, 225). It was therefore reasonable to presume that transmission at this muscle receptor-DSCT cell synapse comprised a high-fidelity relay. This interpretation was further supported by later intracellular recordings showing that stimulation of these same afferents evoked larger amplitude and longer duration excitatory postsynaptic potentials (EPSPs) in DSCT neurons than in motoneurons (see Fig. 1, C and D; Refs. 59, 70, 74-76, 225). Thus most of the earlier research on the DSCT was based on the assumption that the direct, monosynaptic input from sensory receptors was the only functionally relevant input.

For over a decade, electrical activation of afferent input was the primary "functional" means used to examine the connectivity patterns between peripheral nerves and DSCT neurons. Despite the anatomical complexity of Clarke's column shown to contain both large multipolar projection neurons (Fig. 1A) and interneurons (194, 282), most of the electrophysiological research focused on the shortest latency responses to nerve stimulation. These studies led to an elaborate set of divisions and subclasses based on both modality and receptor specificity (201). According to this classification, the DSCT consisted of proprioceptive and exteroceptive subdivisions, each being further parceled into subclasses depending on the main type of afferent input. Thus neurons in the proprioceptive subdivision were shown to receive monosynaptic excitation either from muscle spindles (primary or group Ia afferents, with minor convergence from secondary or group II afferents) or Golgi tendon organ (group Ib) afferents with a high degree of spatial discrimination (70, 200, 203). Another three groups of DSCT neurons were distinguished within the exteroceptive subdivision. One group was activated from touch and pressure receptors in the skin, another from slowly adapting pressure receptors in the footpads, and the third by receptors that elicited flexion reflexes (133, 178, 201, 334).

Although little attention was paid to polysynaptic pathways, it was nevertheless evident from even the earliest studies (191) that single shock electrical stimulation evoked substantial long-latency responses. Because many of these responses were inhibitory, they were viewed as another example of the well-known surround inhibition that was shown to provide a sharpening of point-to-point resolution through contrast enhancement in the visual system (132, 190, 191). This interpretation helped to reinforce the notion of a functional organization based on precise topological projections. In fact, reviewing the current status of research on spinocerebellar systems in 1965, Oscarsson concluded: " ... DSCT [neurons] carry information about proprioceptive and exteroceptive events with high degree of spatial discrimination ... used in the fine coordination of posture and movement of the individual limb" (257).

Figure 2 summarizes some of the main points of this proposed organization. It shows a number of direct monosynaptic connections from specific receptor types (bold arrows) and a few interneurons providing mostly disynaptic inhibition.

View larger version (35K): [in this window] [in a new window]   Fig. 2. The first organizational scheme proposed for the dorsal spinocerebellar tract circuitry proposed by Lundberg and Oscarsson in the 1960s and later reviewed by Mann in 1971 (206). They proposed two major subdivisions: a proprioceptive division (A) and an exteroceptive division (B). This proposed organization resulted from the emphasis placed on monosynaptic connections and was based on the use of electrical stimulation of peripheral nerves and the responses they evoked in DSCT neurons. The sensory receptor origins of the afferents in the muscle nerves stimulated electrically were estimated from the stimulus thresholds. The lowest threshold group Ia afferents were from muscle spindle primary endings, and the group Ib afferents were from Golgi tendon organs. The higher threshold group II afferents were from muscle spindle secondary endings. The inhibitory responses observed with this paradigm were accounted for in the organizational scheme by including inhibitory interneurons to provide sign changes between sensory receptors and the DSCT neurons. Responses to more intense stimuli that also evoke flexion reflex responses were more complex and were accounted for in the proposed circuit by an oligosynaptic pathway from the sensory receptors referred to as flexor reflex afferents (FRA) to the DSCT neurons in the exteroceptive division.

	VI. NATURAL STIMULATION

Top Previous Next References

Later studies departed from the electrical stimulation paradigm by employing more "natural" stimulation such as the stretch of isolated muscles. The aim of these studies was to provide a more functional description of the specific connectivity patterns between sensory afferents and DSCT neurons. However, the focus was still on specific stimuli. For example, work by Jansen and Rudjord and co-workers (153, 156) contributed to a further distinction within the "proprioceptive" division between neurons receiving inputs from only primary or secondary muscle spindle endings. Intracellular responses to muscle activation (Fig. 1C) also allowed an analysis of the functional convergence of afferents from a specific muscle and its effect on DSCT firing (152). The analysis suggested that the convergence increased the fidelity of the information transmission.

Attempts were also made using muscle stretch to characterize the interaction between excitation and inhibition from various muscles onto individual DSCT neurons (151, 154, 155). These studies concentrated on the cellular mechanisms of inhibition and concluded that both presynaptic and postsynaptic mechanisms were involved.

A more in-depth analysis of the cutaneous input to the DSCT was also attempted by Mann (205), who extended the earlier observations by Lundberg and Oscarsson through the use of natural cutaneous stimulation to characterize receptive fields and receptor specificity. He found that only about one-fifth of the recorded neurons were strictly cutaneous, and their receptive fields were not discretely organized, being larger proximally and occasionally being "broken receptive fields." A small fraction of cells received a combination of cutaneous and deep (interpreted as muscle) inputs, and interestingly, the cutaneous and deep receptive fields were often not congruent. Although he did not study the inhibitory effects on DSCT discharges systematically, he did note that cutaneous inhibitory fields had irregular shapes that were usually eccentric to and not surrounding the excitatory fields. He also found that 20% of the DSCT neurons recorded in this study could not be activated by any localized peripheral stimulus (mute cells).

Mann (206) pointed out the potential complexity of DSCT organization emerging from this study again in a later review. Although he did not dispute the classification originally proposed by Lundberg and Oscarsson (see Fig. 2), he did raise concerns about oversimplifications that were made about the functional organization of the DSCT. Mann concluded the review by remarking: "The DSCT has received a good deal of attention, perhaps because of its accessibility and presumed simplicity. The accessibility is real, but its simplicity may be brought into question ... A clear separation of cutaneous and muscle DSCT is no [longer] possible ... Thus it seems useful to think of the DSCT as a mixed tract rather than to invent new terminology for the cutaneous subdivision, confusing an already somewhat muddy taxonomic picture". Mann also questioned the specificity of DSCT neurons originally classified as group Ia and group II: "Coded in the discharge of group Ia and group II DSCT cells is the length of a muscle under stretch, and the coded information has been altered little by the imposition of the synapse ... It is well documented however, that receptors of more than one muscle group can cause a given cell to discharge leading one to question how well the length information for one muscle can be isolated from the DSCT discharge when another muscle is stretched simultaneously."

Although Mann's review had the merit of pointing out limitations in the classical approach and the resulting view of the DSCT, it ultimately failed to provide a clear alternative functional framework. In fact, the traditional thinking about sensory projections was also evident in the interpretations given to results from a pioneering behavioral study by Arshavsky et al. (12) in which they recorded DSCT activity during treadmill locomotion in decerebrate cats. For example, they observed that at least 50% of the DSCT neurons were active in more than one phase of the step cycle, but they attributed this behavior to receptor input from biarticular muscles and did not consider the possibility of convergence mechanisms from multiple hindlimb muscles. Consequently, they concluded that "our findings confirm the view, based on the study of the afferent connexions of DSCT neurones, that the DSCT transmits information about the activity (i.e., the phase and the strength of the contraction) of separate muscles or of a few synergists."

Thus the use of natural stimulation did not lead to a fundamentally different model of spinocerebellar organization because it was still subject to the same methodological limitations and to the same basic assumptions. The focus on isolated stimuli failed to acknowledge an integrative role for the spinal circuitry, so the implied function of the DSCT was to relay a high-fidelity copy of highly localized sensory information to the cerebellum. Any integrative function required for appropriate sensorimotor integration was the exclusive purview of the cerebellum and other central structures.

	VII. POLYSYNAPTIC PATHWAYS

Top Previous Next References

This relay model was challenged when it could be shown that converging polysynaptic pathways play an extensive role in the activity of the DSCT neurons. The issue was revisited through a series of studies that introduced more quantitative analytical techniques employing a sensitive cross-correlation analysis (180, 249) to characterize DSCT responses to electrical stimulation of hindlimb peripheral nerves (179). The results of this analysis showed that most DSCT neurons respond to stimulation of flexor and extensor nerves in the proximal and distal hindlimb. This implied a much larger pattern of convergence than was previously assumed from intracellular recordings, and it contributed to the finding that the most common DSCT responses, accounting for ~80% of the total responses, were polysynaptically mediated. Figure 3 shows examples from a single cell reported in that study illustrating long-duration inhibitory (type 2) and long-latency excitatory (type 3) responses.

View larger version (21K): [in this window] [in a new window]   Fig. 3. Cross-correlation analysis revealed a variety of responses to peripheral nerve stimulation and a greater degree of convergence than had been seen with intracellular recording. Extracellular action potential recordings from a single DSCT neuron in response to electrical stimulation of muscle nerves were used to generate cross-correlograms to quantify the poststimulus time probability density of spike occurrence. These records show the time course of excitability changes from the mean firing level (dashed lines) following single shocks to the sensory axons in four different muscle nerves. They illustrate a convergence of sensory input from receptors in the hamstring and quadriceps muscles proximally and the gastrocnemius-soleus and anterior tibial muscles distally. A: type 1 (monosynaptic) response to hamstring nerve stimulation. B: type 2 (inhibitory) response with long duration evoked by quadriceps nerve stimulation. C: type 3 response (oligosynaptic excitatory) response to anterior tibial nerve stimulation that differs from the type I monosynaptic response by its longer latency and often longer duration. D: a very weak type 2 (inhibitory) response with a short duration evoked by stimulation of the gastrocnemius-soleus muscle nerve. [Redrawn from Knox et al. (179).]

By the early 1980s, the role of interneuronal networks in the function of DSCT neurons had become more apparent. The available evidence for some of the circuitry underlying the polysynaptic organization of spinocerebellar systems was summarized in a review by Bloedel and Courville (28). Their concept of the spinocerebellar circuitry now included both excitatory and inhibitory interneurons interposed between the sensory receptors and the DSCT projection cells as summarized in Figure 4.

View larger version (38K): [in this window] [in a new window]   Fig. 4. An organizational scheme for the dorsal spinocerebellar tract circuitry proposed by Bloedel and Courville (28) to synthesize the evidence available in 1981. In contrast to the earlier proposal illustrated in Figure 2, this proposal details a number of oligosynaptic pathways to account for nonmonosynaptic effects and for the results from experiments utilizing natural stimulation such as muscle stretch in place of electrical stimulation. Except for the added complexity of the circuit diagram, however, the proposal recognized the same basic organization as the earlier proposal with a proprioceptive (A) and cutaneous (B) division. For clarity of presentation, here we omitted from the diagram the monosynaptic convergence from group Ia and group II afferents that is shown in Figure 2 and the presynaptic inhibition of the monosynaptic pathways. [Redrawn from Bloedel and Courville (28).]

The issue of polysynaptic convergence also became the "leit motif" of a long series of experiments by Osborn and Poppele (248-254) that spanned almost a decade. In these studies, DSCT responses to several types of sensory stimulation, from electrical nerve stimulation to passive ankle joint movements, were analyzed with a variety of quantitative techniques to gain insight about the functional significance of the pattern of sensory convergence in the DSCT.

The results of these studies led to several new observations about the organization of the DSCT circuitry. First, they found that muscle contraction is the most potent single stimulus and inhibition the most common response to that stimulus (248). Then, by extrapolating from large population samples, they showed that >60% of the DSCT population representing sensory input from the entire hindlimb responds to stimulation of primary afferents from a single muscle group (the gastrocnemius-soleus), and mostly via polysynaptic pathways. Thus not only did the DSCT neurons appear to be strongly modulated by muscle force, but also "the information carried by the DSCT is not discretely organized muscle by muscle." They went on to observe that "the role of integrating afferent information from the various muscles and joints of the limb no longer seems to be the exclusive province of the cerebellum, but also appears to be shared by spinal centers such as the DSCT" (250).

Although each separate stimulus seems to affect a substantial fraction of the DSCT population, more complex stimuli like joint rotations can affect nearly the entire population (88%, Ref. 46). Thus it became evident that even relatively localized stimuli such as single joint movements have a widespread effect when they activate a variety of receptor types simultaneously (e.g., cutaneous, muscular, and joint receptors). Furthermore, the response patterns evoked by these more comprehensive stimuli were distributed differently across the population than were those evoked by more discrete stimuli. This was evident in the comparison of responses to stretch of the gastrocnemius-soleus with those evoked by passive ankle flexion (254). The two stimuli were designed so they stretched the muscles identically, but they elicited different response components in the population. This result implied that converging input from nonmuscle receptors or from other muscles acting at the ankle joint can modify the activity of DSCT cells so that the same local stimulus, in this case stretch of gastrocnemius-soleus, can evoke a different behavior in another context. It was a clear demonstration that DSCT responses do not merely reflect the activity in specific classes of sensory receptors.

Although the polysynaptic model of spinal proprioception that emerged from these studies seemed more difficult to place in a functional framework than the simpler direct-relay models, it did seem to be more consistent with the anatomy of the system. For example, Walmsley and Nicol (324) argued against a topographical arrangement of muscle input within the DSCT based on a study of muscle projections using intracellular recording and WGA-HRP filling. They found that neurons activated monosynaptically from ankle extensors were scattered throughout Clarke's column, and they generally received converging input from more than one muscle. Along this same line, Zytnicki et al. (339) found a large heterogeneity in the responses evoked in DSCT neurons by contraction of the anterior tibialis muscle. A convergence of cutaneous sensory inputs was also implied by the results of Kim et al. (177), who examined responses to stimuli applied to the skin of the cat footpad.

The increasing evidence that various neuronal subgroups residing outside Clarke's column also contribute axons to the DSCT (119, 210, 211) suggests further that these subpopulations might also have distributed connectivity patterns. Two studies in the late 1980s provided careful examinations of the inputs to these non-Clarke's column neurons. One study characterized the sensory input to DSCT neurons located caudal to Clarke's column in laminae V and VI of the lumbar spinal cord (11). These authors found an extensive mono- and disynaptic pattern of convergence from electrically stimulated muscle as well as cutaneous and joint nerves compatible with a major integrative role for these neurons. A similar experimental approach was used by Edgley and Jankowska (73) to investigate the connectivity of more rostral dorsal horn spinocerebellar neurons. These neurons were found to receive extensive monosynaptic convergence from group II muscle afferents, as well as from joint and cutaneous afferents. Unlike the neurons studied by Aoyama et al. (11), however, they did not seem to receive input from group I muscle afferents.

	VIII. PARALLEL DISTRIBUTED NETWORK

Top Previous Next References

Rather than suggesting a specific functional organization, the large-scale sensory convergence onto DSCT neurons and associated polysynaptic circuitry resembled instead a parallel distributed neural network having a widespread interconnectivity. The analogy to a neural network was reinforced by the results from a set of experiments using a variety of stimuli including muscle stretch or contraction and joint ankle flexion/extension (253). Cells that responded similarly to one of these stimuli generally showed different response profiles for other stimuli. In fact, it was not possible to predict a cell's response to one type of stimulus from its response to another. This argued strongly against any clear-cut functional distinctions among DSCT neurons and led to a proposed scheme in which divergent projections from sensory receptors (input layer) project onto various interneuronal pathways comprising a "hidden layer" in a neuronal network. The weighted pattern of convergence of sensory information from the hidden layer to the individual units of the output layer (the DSCT neurons) determined their firing properties (253; Fig. 5).

View larger version (37K): [in this window] [in a new window]   Fig. 5. An organizational scheme for the DSCT circuitry proposed by Osborn and Poppele (253) as a parallel distributed network. Rather than the discrete circuits proposed in earlier schemes, this proposal used a neural network to represent the diversity of response behavior observed in large populations of DSCT cells. The distinction between cutaneous and proprioceptive subdivisions of DSCT is abandoned in this proposal in favor of a network-like connectivity that included sensory input from muscle, joint, and cutaneous afferents. All of these sensory inputs converge on excitatory and inhibitory circuits that provide variously weighted connections to the DSCT neurons. The experimental results indicated that the excitatory and inhibitory pathways each gave rise to either short latency (early) and longer latency (late) responses. The synaptic input to the DSCT neurons was proposed to come from these circuits as well as from direct monosynaptic projections from receptors. Various weightings of the synaptic connections are illustrated in the diagram by different-sized arrows. The analogy to a neural network was that the sensory receptors acted as an "input layer" having a highly convergent and divergent connectivity with a "hidden layer" circuitry, which in turn has a convergent and divergent connectivity with an "output layer" of DSCT cells. [Modified from Osborn and Poppele (253).]

More generally, the parallel distributed organization proposed for the DSCT circuitry resembles analogous distributed networks that have been proposed for the control of various types of spinal reflexes. The scratch reflex in the turtle is one example in which distributed activity of broadly tuned propriospinal interneurons coordinate two forms of scratch reflex behavior (rostral or pocket scratch) (25, 26). The network controlling this reflex behavior contains shared components with the circuitry devoted to the control of swimming behavior, and it might also be distributed bilaterally (83, 84, 307, 308). Another example is the wiping reflex in the frog (108) in which an irritating stimulus on the frog skin can evoke a precise foot wipe directed at the stimulus using different limb configurations that depend on stimulus location. Stimulating boundary areas between two adjacent wiping zones can elicit the limb configuration of either zone with similar probability, suggesting that a distributed organization might also govern this behavior (171, 172).

A similar distributed network scheme has also been proposed for the control of cat locomotion (274). Although it is well established that some specific receptor types may play quite specific roles in the timing and excitation of locomotion phases (8, 9, 27, 54, 66, 92, 115, 122, 130, 131, 224, 264, 266, 275), the evidence for task-dependent reflex reversals (7, 67, 68, 90, 118, 258, 263, 309; see also Refs. 63, 261, 285, 336 for reviews) and positive force feedback (10, 111, 115, 129, 263, 275, 276) motivated a proposal for a much less specific organization. This organizational scheme developed from the notion of a parliamentary principle originally introduced by Bässler (20, 21), which was an alternative to the idea that distinct interneuronal groups might control distinct phases of the locomotion cycle. It was based on experimental evidence from the stick insect showing that interneurons are all active throughout the phases of locomotion. From this observation, Bässler conceptualized a population-coding scheme for the control of locomotion whereby each interneuron contributed with its level of activity to all the locomotion phases in a manner analogous to the way a member of parliament contributes to a decision with a vote.

The control scheme for mammalian locomotion proposed by Prochaska (274) is similar to this in principle, since it postulates that weighted contributions from various spinal interneuronal network modules might govern locomotion and reflex behavior in general (see Fig. 6). Prochaska (274) attempts to account for reflex variability using the formalism of fuzzy logic, wherein behaviors are selected according to a set of imprecisely defined or "fuzzy" rules relating specific sensory variables to specific motor actions. One basis for such circuit interactions might be the interneurons that are shared among different spinal circuits. In fact, it is likely that a number of interneurons contribute to the control of more than one motor behavior, for example, posture and locomotion (41, 42, 121, 140, 142, 312, 313, or for a detailed review, see Ref. 147). Interneurons may also be shared between reflex and ascending systems. At least one class of inhibitory interneurons that makes synaptic contact with both DSCT neurons and motoneurons (134) may provide for a coupling of afferent information processing between the two pathways. On the other hand, collaterals of primary sensory afferents directed to either DSCT or spinal motoneurons may also be under differential supraspinal (cortical) control via presynaptic inhibition (135, 136, 149, 150, 197). Differential presynaptic control of different sensory afferent branches might, therefore, represent a mechanism by which the activity of the two networks could also be uncoupled. Thus even these relatively simple mechanisms can lead to a number of possible outcomes.

View larger version (36K): [in this window] [in a new window]   Fig. 6. A scheme proposed by Prochazka (274) to account for an interaction between sensory input and motor output during locomotion based on fuzzy logic principles that relate specific sensory variables to specific motor actions according to a set of imprecisely defined or "fuzzy" rules. This scheme proposes a modular processing of sensory inputs to produce a set of muscle activity patterns or motor membership functions. Different weightings of these motor membership functions (indicated by percent values next to the corresponding function) may then produce different behaviors (slow walk or fast run, for example) with the identical combination of sensory afferent inputs. The scheme was motivated by the need to account for task-dependent changes in reflexes, and it resembles an organizational scheme originally proposed by Bässler (20) to explain insect locomotion. In that case, interneurons were found to be active in all phases of locomotion, and the behavior was related to a dynamically weighted sum of their activities. These proposals are both formally similar to the network-like organization depicted in Figure 5. In all cases, there is a large-scale connectivity leading to behaviors that depend on connectivity weightings that may vary dynamically. [From Prochazka (274).]

	IX. INFORMATION ENCODED BY DORSAL SPINOCEREBELLAR TRACT ACTIVITY

Top Previous Next References

The need to adequately engage sensory input was particularly well illustrated for the DSCT circuitry by an experiment in which DSCT responses to bidirectional, single-joint rotations were compared with those for bidirectional perturbations of the whole limb (33). The result was that neurons responding to ankle-only flexion with the same time course of activity showed a variety of response waveforms to ankle-only extension, whereas bidirectional limb movements consistently produced opposite response waveforms for oppositely directed limb movements (see Fig. 7). Thus only when the sensory input from the whole limb was activated in a normal behavioral pattern did the responses of this system show a clear and consistent relationship to the stimulation. Even though the single joint stimuli produced robust and often large-amplitude responses (177, 253, 254), there was no consistent pattern across neurons that related them to the stimulus parameters.

View larger version (23K): [in this window] [in a new window]   Fig. 7. Dorsal spinocerebellar responses to single-joint versus whole limb stimulation. A: poststimulus time histograms from 15 DSCT neurons in response to flexion and extension perturbations of the ankle joint with the other joints fixed. Each trace shows an individual neuronal response, normalized to the maximum and referred to the mean firing rate (horizontal line). The ensemble of neurons was selected based on the similarity of response time courses for ankle flexion. The same 15 neurons gave a variety of responses to ankle extension, indicating therefore no consistent relationship overall to the stimulus. B: responses of 15 neurons to opposite directions of whole limb movement. In this case, perturbations were applied at the hind paw, and all the joints were unrestrained. As in A, cells were selected based on the time course of their responses to flexion. In this case, however, all 15 neurons gave similar responses to extension that were essentially phase-reversed, indicating a consistent relationship across cells when the whole limb was engaged in the movement. [Modified from Bosco and Poppele (33).]

This experiment provided the first clear, direct evidence that the output of the spinal sensory circuitry might be related to whole limb parameters rather than to local parameters like muscle lengths or individual joint movements. Moreover, the same study showed that the activity of DSCT neurons is broadly tuned for limb movement direction and for the position of the hindfoot. These observations set the stage for a completely different functional concept for spinal proprioception by raising the question of the extent to which global versus local limb parameters are represented in the activity of DSCT neurons.

This question was examined systematically by recording DSCT unit activity during passive placement of the cat hindfoot over most of its parasagittal workspace. The response data were related to limb kinematics to determine whether they best represented individual joint angles or more global limb parameters (see Figs. 8A and 9A for definitions). Kinematic parameters that included multiple joint angles always explained a greater fraction of the variance in unit activity than did any single joint representation (38). In fact, linear relationships between unit activity level and the polar coordinates of limb axis length and orientation consistently explained the greatest percentage of variance in unit activity. Examples of this relationship are illustrated in Figure 8, C-F. These three-dimensional plots show a planar relationship between unit firing activity and the limb end-point position given by the coordinates of the limb axis.

View larger version (26K): [in this window] [in a new window]   Fig. 8. Broad tuning of DSCT neurons with respect to limb position. The length and orientation of the hindlimb axis shown in A define the foot position relative to the hip. The stick figures in B illustrate hindlimb postures corresponding to 20 different foot locations in a parasagittal workspace. The data points plotted in C-E represent DSCT unit activity recorded in each position plotted against limb axis length and orientation for the 20 foot positions. In each case, the data points fall within a plane, indicating a linear relationship between cell activity and foot position. The three examples show neurons modulated primarily by changes in limb axis orientation (C), length (D), or both (E). The circular histogram in D represents the directions of foot movement producing the maximal responses (preferred tuning directions) for 79 broadly tuned DSCT neurons. The distribution is significantly quadramodal, with preferred directions clustering along axes (dashed lines) oriented at 13 and 103° with respect to the limb axis orientation. [Redrawn from Bosco and Poppele (34).]

View larger version (21K): [in this window] [in a new window]   Fig. 9. Covariation of cat hindlimb joint angles for passive and active postures. For a given hindlimb posture, the angles at the hip (h), knee (k), and ankle (a) illustrated in A are plotted versus one another in a three-dimensional joint-angle space. Joint-angle measurements made while the hind foot of an anesthetized cat was placed in the 20 positions indicated in Figure 8B all fall within a single plane indicated by the joint-angle covariance plane drawn in B for data pooled from six animals (38). When alert cats were required to maintain stance on a tilted platform, the joint angles also showed a planar covariance as illustrated in C. [Replotted from Lacquaniti and Maioli (187).] Although the joint angles were confined to a linear covariance plane both during active stance and passive limb placement, the planes had different orientations. The covariance plane determined from joint angles measured during active stance was rotated ~90° in the ankle-hip plane with respect to the covariance plane determined for the anesthetized cat. This observation is consistent with a control strategy that may take advantage of biomechanical joint-angle constraints by modulating joint-angle relationships, in this case by reversing the relationship between hip and ankle joints.

Another feature of the DSCT responses to passive limb placement is that the activity of many neurons may also relate to the direction of limb movement as it goes from one position to another. In other words, movement and positional parameters can be represented simultaneously by the activity of individual DSCT neurons, but they are not represented independently. The amplitude of the movement response was found to be linearly dependent on the level of position-related activity (34). This relationship between inputs is reminiscent of the interaction between retinal and eye (or head) position signals first described by Andersen et al. (5) for neurons of the parietal cortex. The amplitude of the retinal receptive field response of those neurons was found to be modulated by the eye or head position. This type of modulation, resulting from a multiplicative type of interaction between the two signals, is commonly referred to as "gain field" modulation. Gain fields have now been demonstrated for several CNS structures involved in oculomotor control (39, 323) and spatial orientation (207, 237, 238, 316). Although the functional interpretation of gain fields is still debated, they appear to represent a neural mechanism by which information about multiple parameters may be compressed and combined in a single unit's activity, and they have also been interpreted as evidence for coordinate transformations (4, 6, 288, 302, 310, 338, for reviews; see Ref. 40).

	X. REFERENCE FRAMES TO FORMALIZE RELATIONSHIPS

Top Previous Next References

One way to formalize the relationship between unit activity and limb kinematics is through the definition of a reference frame, which is a spatial domain in which parameters of interest can vary (305). For example, a primary reference frame for muscle receptors is the spatial domain of the muscle, and the coordinates for effective stimuli in that frame of reference might be the one-dimensional coordinate of muscle length. From the relationship between muscles and joints, we may also define another reference frame for muscle receptors that is joint based. The coordinates in this reference frame would be the relevant joint angles.

	XI. INTERACTION BETWEEN BIOMECHANICAL AND NEURAL FACTORS

Top Previous Next References

To explore this issue further, it is important then to examine how the limb segments and joints interact. The cat hindlimb may be considered as a 3-degree-of-freedom linkage of segments allowing for two-dimensional movement of the end point in a parasagittal plane. This relationship illustrates an important point about limb motor control, namely, the need to deal with extra degrees of freedom. In the cat hindlimbs for example, movements about three joints must be controlled to achieve limb end-point movements that are confined to a two-dimensional workspace. However, there is a tight coupling among the three joint angles in the cat hindlimbs during both passive and active limb movements that effectively reduces the limb degrees of freedom from three to two (38, 188), implying in this case there may be no extra degrees of freedom to control. This reduction is illustrated by the finding that the relationship among the three joint angles of the hindlimb shows a planar or two-dimensional covariation over a large range of limb positions. Thus the data representing limb positions in a three-dimensional joint-angle space fall on a plane rather than being scattered throughout the space as expected if joint angles varied independently (see Fig. 9).

The basis for the joint-angle coupling that leads to this behavior in the passive limb is presumably biomechanical, as indicated by post mortem assessment (38). For example, the biomechanical properties of biarticular muscles that span two joints from origin to insertion as well as more passive structures such as ligaments undoubtedly have a role in coupling forces across joints. During movement there may also be inertial coupling between limb segments as well.

Behavioral studies of posture in cats have suggested that neural control of the limb may actually take advantage of these biomechanical constraints (186-188). Cats trained to maintain stance on a tilting support platform also showed a linear covariation of joint angles during this task (187; Fig. 9C). However, the joint-angle covariance pattern in this case is somewhat different from that described in the passive limb (Fig. 9B). First, the coupling among joint angles is tighter, suggesting that neural control may actually reduce further any independent motion in the individual limb segments. Second, the orientation of the covariance plane is different due mostly to a sign inversion of the relationship between the hip and the ankle angles (compare Fig. 9, B and C). These observations suggest that the limb biomechanics may provide a basis for the planar joint-angle covariance that effectively reduces the limb's degrees of freedom. A reduction in degrees of freedom may then simplify control strategies for maintaining stance, since the problem of mapping sets of joint angles into foot positions (188) can be solved by appropriately adjusting the relationships among limb segments. In this case, it would involve somehow reversing the passive relationship between hip and ankle angles. Sensory input that is also constrained by the biomechanical coupling is more likely to support this type of control strategy based on global limb parameters rather than one that controls joints independently.

Recently intersegmental covariations have also been observed and studied during locomotion (27, 120, 298; see also Ref. 184 for a review). These studies extended the observations made during standing behavior to the kinematics of locomotion and also pointed out that limb axis length and orientation represent kinematic invariance across various types of gait. Lacquaniti et al. (184) proposed from these observations that "the CPG [central pattern generator] may control limb segment motion by encoding the waveform of the elevation angles. In response to these kinematic reference signals, the appropriate muscle synergies would be determined in a subordinate and flexible manner to adapt to the current mechanical constraints."

The biomechanical properties of the limb may also provide a substrate for interjoint coordination at the reflex level (114, 181, 335; see also Ref. 243 for a review). Muscle length-dependent excitatory connections exist between muscles with synergistic biomechanical actions across joints (242, 245), whereas force-dependent, inhibitory pathways link muscles that exert torque in different directions (31, 32, 145, 300). Therefore, the action of these connections could result in a stronger across-joint coordination that complements the biomechanical coupling.

These results suggest that biomechanical interjoint coupling may be utilized and perhaps fine-tuned by the CNS in controlling the limb. Such interplay between controller and the controlled plant could play an important role in the process by which the nervous system copes with the potentially complex multi degree-of-freedom control problem. In essence, purely biomechanical factors establish a reduction in the limb degrees of freedom that is reinforced and modified by spinal reflex pathways that interconnect muscles acting at different joints. Consequently, the role and the computational load of the higher level control may be effectively simplified.

	XII. REFERENCE FRAMES FOR DORSAL SPINOCEREBELLAR TRACT ACTIVITY

Top Previous Next References

The question for the DSCT then is the extent to which the activity of these neurons is referenced to the joint-angle covariance or to more abstract limb parameters such as the limb end-point position. In the passive limb there is a single relationship between limb end point and a linear combination of joint angles and thus a unique limb geometry for each end-point position (38). Therefore, the finding under passive conditions that DSCT cells encode limb length and orientation, or effectively limb end-point position is not surprising. In fact, because of the biomechanical coupling, we could expect this result even if the cells received only a minimal sensory convergence across joints.

However, a number of observations suggested that DSCT activity does reflect a contribution from the spinal circuitry in elaborating a representation of limb axis length and orientation. One is that among the cells that responded independently to all three hindlimb joint angles, those that related similarly to one joint angle could show different relationships with the other two (38). In other words, the pattern of joint-angle convergence expressed by these cells was not fixed to the biomechanical joint covariance. The relationship between DSCT activity and joint angles could be weighted differently across cells as in a network or fuzzy logic type of organization. The other observation was that the cells are broadly tuned with respect to limb axis parameters (33), and the preferred tuning directions are clustered along and perpendicular to an axis roughly coincident with the limb axis (34; see Fig. 8E). This observation provided an indirect indication that the weightings of inputs from various limb segments might be biased toward an explicit representation of the whole limb.

The crucial test of this issue was to alter the joint angles and end-point position separately (35, 37). By constraining the movement at specific joints, sets of limb end-point positions could be imposed that were each associated with a different joint-angle covariance pattern. The responses of DSCT neurons for each position in which the limb was free to move was compared with those obtained with joint constraints. The result was that about one-half the sample cell population studied exhibited the same response to the two different joint-angle covariance patterns across l