Prediction is a fundamental component of motor control. For instance, when catching a baseball it is necessary to predict where the ball will be at a given instant and how much force its impact will generate in order to prepare the hand for the catch. Central to motor prediction is the forward model, which enables prediction of the sensory effects of movement (1). Substantial evidence indicates that humans use forward models to predict the sensory consequences of their own actions (1–6), as well as to predict the dynamics of objects in the external environment (4, 7–11). Furthermore, forward models of object dynamics are necessary to guide visuo-motor coordination tasks, and can even override observed kinematic feedback (12, 13).

Predictive forward modeling becomes even more imperative as the reliability of visual feedback is compromised, for example in dim lighting, or disease states such as Parkinson’s disease (PD) where the visual system may be affected (14). Normally, human subjects account for the degree of sensory uncertainty during motor performance by de-weighting their reliance on sensory feedback when it is ambiguous (1, 6, 15–20), and instead more heavily rely on predictive forward models (6, 18, 20). However, when subjects are unable to use predictive motor control, the motor response no longer anticipates sensory feedback but rather reacts to it in an uncoordinated manner (21–24).

Motor performance in PD, at least in the early stages of the disease, is greatly improved by pharmacotherapy, and l-DOPA remains the gold standard of treatment in PD (25). However, l-DOPA-induced dyskinesias (LIDs) – excessive choreoathetoid involuntary movements – are a relatively common side effect of l-DOPA of which peak-dose LIDs are the most common type (26). Though LID pathophysiology remains unclear, behavioral studies suggest that rather than being a purely motor phenomenon, LIDs may be associated with deficits in sensorimotor control (27–29). For example, dyskinetic subjects have demonstrated increased variation in tracking velocity during a visually guided tracking task that was significantly reduced when visual feedback was withdrawn, suggesting an exaggerated motor response to sensory input (27). As dyskinetic subjects are often unaware of their involuntary movements (30), and have been shown to underestimate the distance their limb has moved (28), a component of sensorimotor deficits associated with LIDs may be attributed to impaired predictive motor control. For instance, a mismatch between predicted movement and actual movement may drive dyskinetic subjects to compensate for the sensory discrepancy with excessive movement. Deficits in predictive motor control are typically seen in diseases of the cerebellum (31–36), yet such deficits have also been demonstrated in PD (37, 38). There is evidence to support cerebellar involvement in LIDs (39), and altered activity and plasticity in the prefrontal cortex in dyskinetic subjects (40, 41) may contribute to altered sensorimotor control in LIDs.

If inadequate predictive motor control is an underlying feature of dyskinetic subjects’ motor performance, then a heightened reliance on sensory feedback should be especially prominent in conditions where healthy subjects rely more heavily on predictive forward models, such as when confronted with ambiguous visual feedback (6). Accordingly, we hypothesized that dyskinetic subjects would demonstrate an impaired ability to de-weight ambiguous visual feedback during a visually guided tracking task. We have purposely chosen a motor adaptation task, whereby subjects had to adapt to changing sensory information. We have recently demonstrated that overall, PD subjects are susceptible to sensory uncertainty during visually-guided tracking (42), but in that study we did not dichotomize dyskinetic and non-dyskinetic subjects. We have since recruited additional PD subjects while employing the same tracking task to assess the reliance on uncertain visual feedback of dyskinetic and non-dyskinetic PD (NDPD) subjects. As previous work has demonstrated linear dynamical system (LDS) models to be a sensitive marker of motor performance in PD (42, 43), here we use LDS models in addition to quantifying tracking error to assess tracking performance. By extracting the decay rate parameter from the LDS models during ambiguous tracking, we quantified subjects’ relative reliance on uncertain visual feedback.

Subjects’ characteristics are shown in Table 1. There were no significant differences between age, UPDRS, disease duration, and LEDD (p > 0.05), though converted daily l-DOPA dosage was significantly higher for the DPD group. Analysis of RMS error between trials revealed that there was a significant decrease in the pooled RMS error between trial set 1 and trial set 2 (p < 0.00001) that stabilized between trial sets 2 and 3 (p = 0.7). The individual groups’ RMS error by trial is shown in Figure 1.

The results of the mixed-model ANOVA tests on RMS error are shown in Tables 2 and 3. When comparing DPD and NDPD subjects, ambiguity and tracking speed were significant independent factors on RMS error, as well as the interaction terms between ambiguity and tracking speed, PD group and drug status. When normal subjects were included in the analysis (Table 3), ambiguity, tracking speed, and group were all significant factors, as well as the interaction terms between ambiguity, speed, group, and subject. Similarly, when determining the effects on log(decay rate), ambiguity and speed were significant when comparing DPD and NDPD (Table 4), as well as the interaction between l-DOPA and PD group (Table 4). When control subjects were included in the analysis, significances were seen in the main effects of ambiguity, speed, and group, as well as the interaction effects of ambiguity and speed and group (Table 5).

The differences in RMS error between non-ambiguous and maximum ambiguous tracking conditions were not significantly correlated with UPDRS scores at either tracking speed for either dyskinetic subjects or for non-dyskinetic subjects and p > 0.05.

The overall effect of increasing ambiguity and speed on overall tracking performance, and the l-DOPA effect, is illustrated in Figure 3. As expected, there were increases in RMS error with both speed and visual ambiguity. DPD subjects had the highest error, which was partially ameliorated by l-DOPA. In NDPD subjects, after medication, the tracking error approached that of control subjects.

The effect of visual ambiguity on the log of decay rate is shown in Figure 4. At slow speeds and higher levels of ambiguity, NDPD subjects had lower log(decay rates) than controls (left panel). However, at higher tracking speeds the NDPD subjects had similar or higher log(decay rates) as controls. In contrast, DPD subjects had higher values for log(decay rate) at high ambiguity levels at both speeds, a situation not ameliorated by medication.

The regression analysis illustrates the relative contribution of increasing ambiguity and speed to RMS error (Figure 5) and log(decay rate) (Figure S1 in Supplementary Material) by group. The speed and ambiguity regression coefficients captured by the model were highly significant for all groups (p < 10–5), and the between group differences in both speed and ambiguity regression coefficients were also highly significant (p < 10–5). Figure 5 demonstrates that increasing tracking speed and ambiguity contributed to the RMS error of DPD OFF subjects significantly more than for all other groups. Additionally, the susceptibility to speed and visual ambiguity is not normalized with medication for DPD ON subjects, but is roughly normalized for NDPD ON subjects. Figure S1 in Supplementary Material suggests that log(decay rate) is significantly affected by visual ambiguity in PD, but especially so for DPD subjects. l-DOPA had less of an effect on the sensitivity of log(decay rate) to tracking speed in DPD compared to NDPD subjects.

The Akaike’s final prediction error (FPE) and AIC used to assess the LDS models from ambiguous tracking conditions revealed robust tracking models. The means and standard deviations of the estimated LDS models’ FPE and AIC scores were ≤3.1 ± 2.0 and ≤1.8 ± 0.4 respectively, for all groups across all conditions, which is indicative of high model quality/fit. Furthermore, there were few outliers in the FPE and AIC values indicating validity of the modeling framework across subject groups.

In order to get an intuitive interpretation of the significantly different decay rates, we interrogated typical models from each group (i.e., models with eigenvalues close to the mean for each group) with one-dimensional sinusoidal inputs and additive noise similar to the experiment to determine the predicted tracking performance. Ideal tracking performance would occur in systems that ignore the noisy input and faithfully maintain sinusoidal tracking. Consistent with the RMS error results, the sinusoidal tracking improved with post-medication models in both dyskinetic and non-dyskinetic subjects. However, consistent with the statistical results, the dyskinetic model had an impaired ability to ignore the noisy visual cue, and was excessively reliant on noisy ambiguous visual feedback (Figure 6).

We observed the significant differences in the decomposition of the finger tracking data, depending upon whether or not jitter was present (Figure S2 in Supplementary Material). For only the DPD (D-pre and D-post) group, we found that contribution of FD→(t) – the vector from finger position to the desired position – increased (p = 0.036 and p = 0.026), while the contribution of D→(t) – the desired trajectory – decreased when jitter was present (p = 0.002 and p = 0.005).

We examined the ability of 9 dyskinetic and 10 NDPD subjects, as well as that of 10 age-matched control subjects, to de-weight uncertain sensory feedback and instead rely more heavily on predictive motor control during visually guided tracking. The relative contributions of increasing target ambiguity and speed (examined independently) to the RMS tracking error were the greatest for DPD subjects off medication (Figure 3). As expected, l-DOPA medication improved overall tracking performance for both PD groups, as evidenced by reduced RMS error with medication (Figure 3).

Dyskinetic subjects’ gross motor performance was worse in the Parkinsonian state than in the “on” medication state despite the presence of LIDs. This is consistent with the presence of mild LIDs experienced by DPD subjects that did not interfere with the completion of motor tasks, and the increased movement in the on-state compared to the Parkinsonian state that enabled improved overall motor performance. However in dyskinetic subjects, improved overall tracking performance came at a price: they were also more responsive to and reliant on non-informative visual cueing (Figure 6, right lower panel). We interpret our results in the context of established performance trade-offs in control theory, in which controllers that produce exceptionally fast, high-performance tracking under ideal circumstances are also extremely poor at disturbance rejection (that is, they experience high sensitivity to external or un-modeled noise processes). In a biological context, this is indicative of a system that relies more heavily on ambiguous sensory feedback and less on predictive motor control. This interpretation is further supported by the decomposition of the finger velocity F→(t). For DPD group only (i.e., D-pre and D-post), the contribution of the desired trajectory D→(t), to finger velocity, F→(t), was significantly reduced during jitter cases, implying that finger movement velocity was significantly degraded by jitter in these subjects. Interestingly, FD(t), i.e., where their finger was to where it should have been, significantly increased in the DPD group only, possibly reflecting an compensatory corrective motor movement after the realization that they had been misled by the jitter. The trends we observed in the LDS are consistent with the trends we observed through the regression analysis. The LDS model separates the contribution of the output due to the desired trajectory from the contribution of the output due to the additive jitter. The decreased reliance on D→(t) in DPD subjects is consistent with the increased decay rate in the LDS – that is, the subjects are following the noise more than they were in the no-noise case.

It is important to note that the LDS models utilized here are deterministic. The numerical algorithm used to identify the LDS model, in fact, minimizes the residual between the actual output and the predicted output, in essence capturing as much information as possible from the input–output relationship and leaving only white noise. What would we see if the only differences between DPD and NDPD were that the DPD subjects had the same tracking performance but superimposed dyskinesias that were truly random fluctuating movements? The parameters of the models would be the same, but the residuals of the model, which reflect the part of the movement not accurately captured by the deterministic model, would be much higher in the DPD case. Yet we observed the opposite: the model residuals were not significantly higher in the DPD subjects (as reflected by the lack of significant differences in their model scores) and the parameters of the model were appreciably different. In fact, this raises an important issue, that a key interpretation of our findings is that LIDs may not be “random” at all as is normally assumed, but a deterministic response to various endogenous and exogenous stimuli that is normally appropriately de-weighted. This may explain why increasing sensitivity to stimuli such as that seen in anxiety (53), or increased vigilance due to cognitive or motor load (54) may increase dyskinesias.

Normally, forward models are used to predict sensory feedback, and the predicted feedback is subsequently compared to actual feedback when it becomes available after an inherent delay (55, 56). The difference between the actual and predicted sensory feedback is known as the sensory discrepancy or error, which is then used to update the forward model and in turn improve motor performance (1, 56). In fact, the concept of forward modeling has been extended from predicting the sensory consequences of movement to predicting the external environment (4, 7–13). For example, evidence indicates that human subjects utilize forward models of visual cues (11), of target motion during interception tasks (12, 13), and of the physical laws of gravitational acceleration (4, 9, 15). We quantified RMS tracking error as the difference between the subjects’ index finger position and the target position along the smooth Lissajous path at any given time point. During the non-ambiguous conditions of our tracking task, the sensory discrepancy is likely minimal as the predicted sensory feedback relating the subjects’ index finger position and the target position would be congruent, which is supported by the lack of differences in RMS tracking error between the groups in the baseline conditions (Figure 3). However, in the ambiguous tracking conditions the sensory discrepancy would be large due to the ambiguous jitter of the target. Human subjects have been shown to reliably predict the mean perturbation delivered from a variable distribution in reaching tasks (50, 57), and to do so according to Bayesian inference (17).

Though we did not explicitly test the use of Bayesian statistics in this study, the strategy of more heavily weighting the mean jitter amplitude and de-weighting the instantaneous uncertain jittering position of the target in order to predict the desired tracking position, corresponds to the optimal motor response in our task that minimizes RMS tracking error.

As expected, in our study, normal controls had the lowest overall tracking error of all groups (Figure 3). However, Figure 4 provides interesting insight into how this is achieved. During slow tracking and in high ambiguous situations, NDPD subjects had even lower decay rates than controls, suggesting that they were robust to the ambiguity – so much so that they excessively de-weighted the (still partially meaningful) visual information. However, in the high speed tracking condition, it becomes more critical to de-weight the visual information and the originally sluggish approach of the NDPD becomes the appropriate response – this is why NDPD patients ON medication and controls had essentially the same decay rates. These observations are consistent with other studies demonstrating that NDPD subjects do not overly respond to visual feedback (58), and that healthy human subjects internally account for sensory uncertainty and de-weight uncertain feedback during motor performance (1, 18, 20). In contrast, in both slow and fast tracking conditions, the DPD subjects demonstrated excessively high decay rates (Figure 4), implying faster dynamics, even though this resulted in excessive overall tracking error (Figure 2).

Thus, the inability of DPD subjects to de-weight ambiguous visual data that we observed may be based on excessive sensitivity to discrepancies between a (accurate) forward model and sensory feedback and/or an impaired forward model. The effect of l-DOPA medication may provide insight on this. In addition to reducing overall tracking error (Figure 3), l-DOPA medication made overall tracking error less susceptible to tracking speed and visual ambiguity (Figure 5), but had minimal effect on the log(decay rate)’s sensitivity to tracking speed and visual ambiguity (Figure S1 in Supplementary Material). If we assume that decay rate is related to corrective sub-movements and hence responses to discrepancy between models, this would imply that l-DOPA largely improves forward model accuracy (and hence reduced RMS error’s sensitivity to visual ambiguity and speed) while having minimal effect on the dynamics of the response to the errors between the forward model and sensory estimates (Figure 5).

A possible functional neuroanatomical correlate of the inability of DPD subjects to de-weight ambiguous visual feedback demonstrated in the present study is inadequate predictive cerebellar forward modeling. There is growing evidence of functional cerebellar changes occurring in PD (59–64) and in DPD (39) that supports this possibility. Furthermore, the cerebellum is known to have an integral role in predictive motor control, and predictive deficits that lead subjects to excessively respond to feedback are typically seen in cerebellar disease (21, 24, 31–34). Extensive evidence supports the use of forward models in human subjects (2–6), and neuroimaging, electrophysiology, and transcranial magnetic stimulation (TMS) studies provide strong evidence for the role of the cerebellum in forward modeling (31, 65–75). Interestingly, evidence from neuroimaging studies demonstrates significantly increased cerebellar activity in conditions of mismatch between predicted and actual feedback (66), and the degree of mismatch imposed by temporal delays has been correlated with cerebellar activity (76). Further evidence indicates that the cerebellar climbing fiber-Purkinje cell synapse may signal the error between the predicted and actual sensory feedback (71, 72, 77–80).

In addition to the cerebellum, the posterior parietal cortex (PPC) is believed to have an important role in predictive motor control (65). The role of the PPC in making on-line corrections (a process that requires forward models) during movement has been demonstrated in patients with lesions to this area and through the use of TMS (81, 82). TMS applied to the PPC of healthy human subjects prevented them from making fast on-line corrective movements to a target perturbation in a reaching task when vision of their arm was occluded, and they instead continued to reach to the initial target (81). As DPD subjects in our study were found to be overly responsive to the ambiguous visual feedback (as opposed to unresponsive PPC subjects), this may argue against altered PPC function explaining our results. Moreover, PPC stimulation has been related to motor awareness (83), and interestingly DPD patients can be unaware of their involuntary movements (30). Nonetheless it is possible that altered PPC activity contributed to the impaired predictive motor control of dyskinetic subjects, and as the PPC and cerebellum have reciprocal neuroanatomical connections (84, 85), it is likely that these two structures work together in using forward models to guide motor performance. Given that frontal “executive” dysfunction is well described in PD (86), it is tempting to speculate whether or not impaired frontal lobe dysfunction may contribute to the our observation of impaired forward models in PD. While this may explain, at least in part, the differences between PD subjects as a whole and controls, we do not believe that it could explain the differences between NDPD and DPD subjects we observed.

There is increasing evidence that although dyskinesias are present when DPD subjects are on medications, changes in motor function persist off medication (42). For example, Figure 3 demonstrates that DPD subjects OFF medication were significantly worse in overall tracking compared to NDPD subjects OFF medication. Animal models of PD suggest that unnatural pulsatile stimulation of dopaminergic receptors, occurring with intermittent dosing of l-DOPA, may induce plastic changes that contribute to the development of LIDs (87, 88). Interestingly, younger patients are more prone to developing LIDs (89), and this may be related to a greater degree of plasticity occurring in the younger brain. Additionally, neurochemical changes related to LIDs (90) are not limited to the basal ganglia. Nimura and colleagues (91) demonstrated that the binding potential of the cerebellar sigma receptors was positively correlated with LID scores but not with disease severity of PD patients undergoing pallidotomies; while Koch and colleagues (39) have demonstrated altered cerebellar plasticity in DPD subjects using TMS. Furthermore, we have found behavioral differences that differentiate dyskinetic from non-dyskinetic subjects in the off medication state that may be related to altered cerebellar functioning (42). Thus, the dyskinetic brain may exhibit altered cerebellar plasticity that manifests functionally as inadequate predictive motor control. Direct neuroanatomical pathways connecting the basal ganglia and the cerebellum have been found in primates (92, 93), providing a direct route for the administration of l-DOPA to interact with altered cerebellar structures.

There are a number of potential limitations to our study. First, there was a trend toward greater disease severity of dyskinetic than NDPD subjects, though the difference was non-significant (Table 1). Nonetheless, in order to fully address this, we examined the relationship between UPDRS and the increase in RMS between the non-ambiguous and maximum ambiguous tracking conditions, and no significant correlation for either DPD or NPDP subjects was found. Thus, the worsening of motor performance with increasing visual ambiguity was not associated with disease severity. Furthermore, UPDRS was not significantly correlated with decay rate for either PD groups, except for in the slow tracking condition (ambiguity level = 0.03), which was the only ambiguous tracking condition that lacked significant differences in mean decay rate between groups. Second, in addition to testing while on medication, we also tested PD subjects in the practically defined off medication state with 12 h of l-DOPA withdrawal and 18 h for dopamine agonists, and subjects were symptomatic upon study commencement. We note that this method of examining the practically defined off medication state in PD is universally utilized (94, 95), though we acknowledge that this may not reflect a truly depleted dopaminergic state. Furthermore, non-motor complications can occur with the off medication state (96), as well as pain that can be experienced by DPD patients (97), and such non-motor complications were not accounted for in this study. However, none of the subjects complained of pain and none of the subjects experienced non-motor complications requiring them to stop the study. Third, dyskinetic subjects can experience postural instability while on medication and experiencing LIDs (98), and we did not quantify postural instability between groups. However none of the subjects from either PD group complained of postural problems, and the lack of difference in overall accuracy in the baseline tracking condition of our task indicates that PD subjects were able to perform the task while standing equally well as healthy control subjects, and suggests any potential differences in postural instability did not significantly affect motor performance. Fourth, we did not examine potential differences in visual acuity between dyskinetic, non-dyskinetic, and control subjects. However, once again a lack of difference in RMS error in the baseline non-ambiguous conditions of the task suggests that any differences in visual acuity were not great enough to impact baseline motor performance. Fifth, it is theoretically possible that l-DOPA affected DPD and NDPD subjects with respect to eye movements. However, we note that (99) we found no changes in smooth pursuit gain during dose-related on–off fluctuations, so believe that this was not a factor here.

In conclusion, we demonstrate that DPD subjects are significantly more susceptible to visually ambiguous sensory input during a visually guided tracking task, and that the improvement in overall tracking performance with l-DOPA medication comes at a price for DPD subjects: an increased reliance on ambiguous visual feedback. The results indicate inadequate weighting of predictive motor control in DPD, which may be a significant contributor to pathophysiology of LIDs. We discuss possible cerebellar dysfunction in DPD as a neuroanatomical substrate of inadequate weighting of predictive motor control.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The Supplementary Material for this article can be found online at http://www.frontiersin.org/Journal/10.3389/fneur.2014.00008/abstract