State Estimation in the Cerebellum
- 789 Downloads
An exciting hypothesis about the cerebellum is that its role is one of state estimation—a process that combines efferent copies of motor commands with afferent sensory signals to produce a representation of the current status of the peripheral motor system. Sensory inputs alone cannot provide a perfect state signal because of inevitable delays in their afferent pathways. We have recently reported the effects of transcranial magnetic stimulation (TMS) over the ipsilateral cerebellum as healthy subjects made rapid reaching movements towards visually defined targets (Miall et al. in PLoS Biology 5:2733–2744, 2007). Errors in the initial direction and in the final finger position of this reach-to-target movement were consistent with the reaching movements being planned and initiated from an estimated hand position that was about 138 ms out of date. This interval is consistent with estimates of the delays in sensory motor pathways that would inform the central nervous system of the peripheral status. We now report new data using the same paradigm, testing the effects of varying the TMS stimulus train from one, two, or three pulses. We show that the errors in movement are relatively insensitive to the TMS pulse-train duration. The estimated time interval by which the hand position is mislocalized varied by only 12 ms as the TMS train duration increased by 100 ms. Thus, this interval is likely to reflect physiological processes within the cerebellum rather than the TMS-stimulus duration. This new evidence supports our earlier claim that the cerebellum is responsible for predictively updating a central state estimate over an interval of about 120–140 ms. Dysfunction of the cerebellum, whether through disease or experimental procedures, leads to motor errors consistent with a loss of knowledge of the true state of the motor system.
KeywordsCerebellum Human movement Sensorimotor control State estimation Transcranial magnetic stimulation
Recent work in computational or theoretical motor control has been based around the concepts of internal models , optimal control , and probability . In essence, this broad literature seeks to understand how the central nervous system controls a highly complex nonlinear and often unstable biomechanical system comprising the joints and muscles of the body. It must use optimal—or at least sufficient—strategies in the face of uncertainty about the exact properties of the system due to noise and delay in sensory systems, in central neural pathways, and in motor execution. And, it is assumed that the central nervous system (CNS) maintains information in the form of internal models—neural representations of the properties of the system it is controlling. The cerebellum is thought to be important for learning and storage of these internals models [4, 5, 6], and much experimental research has addressed this hypothesis. A particularly powerful idea is that the cerebellum forms a “forward model”—an internal model that generates a prediction of the expected outcome of motor commands . This predictive signal could be used to plan actions, to help control on-going movements, to coordinate two or more simultaneous actions, and to update a central representation of the current state of the body. It is this last hypothetical role that we address in this paper.
The CNS needs an internal “state estimate” because it does not know the current state of the peripheral motor system because of unavoidable conduction delays in efferent and afferent pathways. Hence, sensory afferents are out of date by the time they reach the CNS, while any recent motor commands may not have yet affected the musculature . Sensory-motor delays vary across modality but may be in the order of 100–300 ms [9, 10, 11, 12]. Thus, the brain generates an estimate of the true state of the peripheral motor system, integrating incoming sensory information with predictions of the consequences of outgoing motor commands [5, 13, 14, 15]. Hence, state estimation is closely coupled with forward modeling . The cerebellum receives ascending proprioceptive inputs and efferent copies of descending motor commands, and it outputs to cortical and brain stem motor nuclei . It also has the necessary adaptive mechanisms to support this hypothesized role, as forward model predictions must be refined and maintained by experience-based motor learning [4, 7]. So, to test the role of the cerebellum in these processes, we have recorded the response properties of Purkinje cells in the lateral cerebellum of monkeys performing a visually guided tracking task and have shown that they encode the direction of movement of a cursor on the screen even when the monkey’s hand moves in the opposite direction . Thus, they appear to code the visual consequences of the actions, as predicted by the forward model hypothesis. Others have reported similar results suggesting the lateral cerebellum encodes the future kinematics of arm movements and not dynamics [17, 18]. The responses of the human cerebellum measured by functional magnetic resonance imaging also confirm its forward model role in coordinating actions, apparently using predictive information from one effector to help control the motion of another [19, 20]. Of course, the cerebellum is interconnected with other brain systems, and the forward model and state estimation processes may be shared with cortical structures including the superior parietal cortex [11, 21, 22, 23] or elsewhere, depending on task .
In a recent paper , we reported the effects of brief disruption of the cerebellum during a reach-to-target action. We used transcranial magnetic stimulation (TMS) over the ipsilateral cerebellum during subject’s reaction time to respond to an auditory cue and to point, without vision, to a previously observed target location. We predicted that TMS-induced disruption of the cerebellum would lead to inaccuracies in movement because the state estimation signals would be lost. A rapid reaching action made without accurate knowledge of the current state of the hand (its position and velocity) would overshoot , analogous to the hypermetria of cerebellar patients . Moreover, if the reaching action started while the hand was already in motion, the initial direction of the movement towards the target would be inaccurate. For example, during a rightwards movement, plans to move forward to the target should reflect the continuously updated estimate of the hand’s current position. Without an accurate state estimate, the reach-to-target would be misdirected, reflecting the mislocated starting position of the hand. Both these predictions were borne out—there were small but significant increases in terminal errors when 100 ms of TMS was applied over the cerebellum but not elsewhere, and cerebellar TMS induced movement direction errors when the reaching movement started during a rightwards or leftwards action but not if the hand was initially stationary . This last result reflects the expected insensitivity of movement direction if the starting posture is known and fixed, as the current state estimate is then no longer time dependent.
However, this work could not resolve an unknown relationship between the fixed duration of the TMS stimulation that we used and our estimate of the functional effect of about 138 ms. It was not clear if the 138 ms merely reflected the TMS duration (100 ms) plus an additional neuronal recovery time (38 ms).
In the present work, we directly test this relationship by repeating the original experiment but varying the TMS stimulus train from one to three pulses at 50 ms intervals. In other words, we varied stimulus duration from 0, 50 to 100 ms and tested its effects on terminal errors, on directional errors, and on the duration of the functional effect. We hypothesize that the state-estimation interval is unrelated to the TMS duration and instead reflects the interval over which the cerebellum predictively estimates the current state of the motor apparatus. Our null hypothesis is that the interval is directly related to the duration of the TMS-induced disruption of the cerebellum.
We predict effects on azimuth angles in the dynamic task , and so we first screened participants for an effect of TMS on this measure with a very liberal inclusion criterion of p < 0.5. We rejected data from three participants who showed no differences between TMS and non-TMS trials, probably because of inter-subject differences in stimulation threshold or in coil-to-cerebellar distance (the coil was positioned relative to the inion). The remaining ten data sets were then analyzed in detail. Group mean trajectories for TMS and non-TMS trials in the dynamic task are shown in Fig. 1b. We concentrate in these results only on the dynamic task results. There were no significant differences in terminal errors between the static and dynamic task (analysis of variance, F(1,6) = 0.35, NS) or for elevation angles (F(1,6) = 0.24, NS); nor were the condition by TMS interactions significant. Expected differences in azimuth angle between the two conditions  are described below.
Directional Error: Azimuth Angle
For the dynamic task, measurement of the difference in average direction between the TMS and non-TMS trials showed a significant TMS-induced change in azimuth angles for all three TMS conditions (Fig. 2b; t(9) > 4.8, p < 0.001) and, again, no significant difference between the new and old three-pulse conditions (t(40) = 0.78, p = 0.437). As expected, there were highly significant differences between the dynamic and static conditions (F(1,6) = 20, p = 0.004), as azimuth angles were not significantly different from zero in the static condition and did not change significantly with increasing TMS number. The interaction between task condition and TMS number was not significant (F(2,12) = 2.76, p = 0.10).
There were also significant changes in elevation angles (Fig. 2c; t(9) > 2.7, p < 0.02) and no difference between the new and old groups (t(40) = 0.82, p = 0.417). There was a trend towards an increase in the angular deviation as the number of TMS pulses increased, although this was not statistically significant in azimuth (F(2,18) = 0.398, p = 0.68). For elevation, this was significant (F(2,18) = 4.39, p = 0.028), with a linear increase in elevation error with increasing TMS pulses (F(1,9) = 13.1, p = 0.006).
State Estimation Interval
These results confirm that TMS over the lateral cerebellum during dynamic actions resulted in a directional deviation of the reaching movement and in increased positional error. These results confirm previous findings that the effects depend on the task, and TMS in the static condition does not generate significant errors in azimuth. Moreover, they strongly argue that the magnitude of the behavioral effect observed is not simply a reflection of the duration of the TMS train, as there was only a limited sensitivity to one, two, or three TMS pulses. Of particular importance is the finding that the estimated time by which the hand is mislocalized (Fig. 3) varied only 12 ms as the TMS stimulus train increased by 100 ms.
It is not certain what duration of effect a single TMS pulse would have on processing within the cerebellar system. Judging by its influence on primary motor cortex, the effect could be brief: there is a narrow window of 5–10 ms within which motor cortical excitability is reduced after each cerebellar shock [27, 28]. However, this indirect measure of cerebello-thalamic-cortical modulation does not indicate how long the cerebellar cortex itself takes to recover, and it is likely to be much longer. Judging by the effects of TMS over visual cortex on image detection  or over motor areas on reaction times , neocortical recovery takes 30–60 ms after a supra-threshold pulse. Adding ∼30 ms to the one-, two-, or three-pulse trains, we might expect our TMS protocol to disrupt the cerebellar hemisphere for about 30, 80 and 130 ms, respectively. The latter is close to the observed interval (Fig. 3). But, it is also close to the delay of sensory afferents informing the CNS about the previous hand positions [9, 10, 15]. And, since the differences between one, two, and three pulses were not significant (except in elevation angle), we suggest that the original 138 ms state estimation interval, based on a large sample size, is instead a good measure of the time over which the cerebellum anticipates the state of the ongoing hand movements.
There was, however, a trend for an increased effect with longer TMS trains, and the elevation angles were significantly related to the number of TMS pulses. This may indicate temporal summation of the effect, such that the three-pulse stimulus is more effective in disrupting the cerebellum than a single pulse . Given that even three-pulse TMS does not completely block processing in the relevant arm areas of the anterior and/or posterior cerebellum, our estimate of ∼120–140 ms may be a slight underestimate but is not likely to be badly wrong.
Another issue is the localization of these TMS effects in the cerebellum. Prior research using TMS to condition the excitability of motor cortex [27, 28] is consistent with the brief excitation of the Purkinje cells, leading to cerebellar inhibition. It is also clear that the target site affects the hand area of motor cortex, consistent with activation of lateral cerebellar cortex, probably in lobules V and VI. It is not clear how widespread within neighboring areas of the cerebellum the effect is.
It also remains for us to test whether the effects we report depend only on the cerebellum or whether the state estimation process is shared with other areas  and whether the cerebellar state estimates are used to update other cortical regions. These will be the aims of further experiments.
This work was supported by the Wellcome Trust.
This article is distributed under the terms of the Creative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.
- 4.Miall RC, Weir DJ, Wolpert DM, Stein JF (1993) Is the cerebellum a Smith Predictor? J Motor Behav 25:203–216Google Scholar
- 6.Paulin MG (1989) A Kalman filter theory of the cerebellum. In: Arbib MA, Amari S (eds) Dynamic interactions in neural networks: models and data. Springer, New YorkGoogle Scholar