Focused Review on Neural Correlates of Different Types of Motor Errors and Related Terminological Issues

Errors in motor actions are thought to be discrepancies between an intended state and an actual state, respectively an estimate of the latter. That such a discrepancy is detected and evaluated as an error is crucial for adaptive behavior and learning. In this regard, it is interesting to investigate the time scales of error processing (when does error valuation happen?). In the last three decades, considerable research has been conducted on error processing. Neurophysiological studies in particular have sought to provide evidence for mechanisms of error processing that monitor the actual state, detect deviations from the intended state, and initiate error correction or compensation. In 1991 (Falkenstein et al., 1991) and 1993 (Gehring et al., 1993), two research teams independently reported an event-related brain potential (ERP) correlating with error responses in choice reaction time experiments. This error negativity (Ne) or error-related negativity (ERN) is a negatively deflecting electroencephalogram (EEG) signal with fronto-central scalp topography peaking 50-100 ms after the erroneous response (dependent on whether response onset is measured by button press or electromyogram; Krigolson, 2018). The functional significance of the Ne/ERN has been explained by different theoretical accounts, which will not be elaborated here (for an overview see Hoffmann and Beste, 2015), but, there is increasing evidence for the reinforcement learning hypothesis (Holroyd and Coles, 2002) that interprets the Ne/ERN as an indicator that an action outcome is worse than expected. It is postulated that the medial frontal cortex (MFC), specifically the anterior cingulate cortex (ACC), receives an error signal from the mesencephalic dopamine system, which is then used as a training signal to adjust subsequent behavior. In addition to the Ne/ERN, a similar ERP component has been found that correlates with external error feedback. This feedback-ERN or FRN peaks about 230–350 ms after the presentation of a negative action outcome (Miltner et al., 1997). There is agreement that the Ne/ERN and the FRN share the same neural source (Miltner et al., 1997; Holroyd and Coles, 2002; Holroyd et al., 2004; Gentsch et al., 2009; Walsh and Anderson, 2012), but show a clear difference in temporal location. The Ne/ERN is time-locked to the action (or response), and emerges shortly after movement onset and prior to any external feedback about the action outcome (see Fig. 1). The FRN, in contrast, is time-locked to the external outcome feedback, and is elicited after its presentation (see Fig. 1). Consequently, the reinforcement learning account posits that the Ne/ERN reflects processes of error valuation based on internal predictions (using a copy of the efferent signals sent to the muscles). While the FRN presumably represents a postdictive error signal based on the integration of external sensory error information (Holroyd and Coles, 2002; Walsh and Anderson, 2012; Krigolson, 2018).

Figure 1

The Ne/ERN and the FRN have increasingly been investigated in more complex motor tasks (e.g., Krigolson and Holroyd, 2006, 2007a, 2007b; Krigolson et al., 2008; Anguera et al., 2009; Vocat et al., 2011; Torrecillos et al., 2014) as opposed to the hitherto used stimulus response tasks. However, in complex motor tasks, an error is not limited to be made on a higher cognitive level by a wrong action selection (e.g., pressing the correct or the wrong key). Errors can occur as well in the way the action is executed (even though an appropriate action might have been selected). Putting it differently, error information could either relate to the movement itself or the action effect produced by that movement. This has consequences for the valuation of an error. A mismatch can be detected between intended and actual movement parameters (such as movement kinematics or dynamics), or between intended and actual movement outcomes. The former might be correctable closed-loop (depending on the velocity of movement execution), while error valuation in the latter is the basis for a correction in subsequent trials. Hence, it becomes necessary to distinctively define the term error or to determine what part of the action is used as a source for error valuation, and to distinguish the underlying neural correlates with respect to their functions. Only with an unambiguous assignment of neural correlate and function, can the correlate be used to investigate processes of error valuation and learning. For instance, action outcome predictions are postulated to positively affect learning (Jordan and Rumelhart, 1992): The better the prediction, the better the error attribution and valuation and the better the learning. These interrelations could be examined by the neural correlate of error prediction (i.e., the Ne/ERN; Maurer et al., 2015), but only if there is certainty about the observed neural processes to clearly indicate error prediction processes.

The present article has the rationale to give an overview of example studies investigating ERPs of error valuation in motor tasks, and to classify them with respect to movement errors or action effect errors. Further, we argue that, depending on the motor task, movement and action effect, errors can coincide or can be distinct, which affects the terminology used for and the interpretation of neural effects, as we will elaborate in the following. We want to draw attention to these differences and sensitize scientists to the use of precise definitions and accurate terminology when referring to neural correlates.

Empirical findings presented to date indicate that error information is processed throughout a complete movement, starting in the planning phase, continuing during monitoring of the execution, and finally being included in the result evaluation (Desmurget and Grafton, 2000; Ridderinkhof et al., 2004). Different types of information are used for this purpose. They differ with respect to the sensory modality (e.g., visual, auditory, kinesthetic), the source of the phenomena/events (e.g., internal [to the body] vs. external events), and the parameters being described (e.g., kinematics, forces, torques etc.). For a button press movement or a reaching movement, movement execution (e.g., position of end effector) and action effect (position of end effector relative to a target position) can be described in an almost identical way. However, in more complex motor tasks, like targeted throwing, the movement (e.g., described by release angle and release velocity) is linked to the outcome (e.g., distance to the target) by a redundant multi-variate non-linear function imposed by the physics of the natural environment. Since error sizes in movement parameters do not necessarily map to equivalent error sizes in the outcome, these conceptual differences add to the differences observed empirically in the dissociation of movement related “low-level” errors, and action-effect related “high-level” errors (Krigolson and Holroyd, 2006, 2007a, 2007b). High-level errors indicate that an action goal (e.g., reaching a target) has been missed or will be missed. These errors correlate with fronto-central potentials with negative polarity (de Bruijn et al., 2003; Krigolson and Holroyd, 2006, 2007a, 2007b; Krigolson et al., 2008). In contrast, low-level errors in movement execution (e.g., a deviation from a designated movement trajectory) are mostly connected to activity in the posterior parietal cortex (Desmurget et al., 1999; Desmurget and Grafton, 2000; Krigolson and Holroyd, 2006, 2007b; Krigolson et al., 2008). Importantly, these trajectory discrepancies can be corrected in principal, such that the action outcome is still achieved (Krigolson and Holroyd, 2007b). However, in cases where a correction fails, the action goal will be missed and, thus, the low-level movement error leads to a high-level error. Since the valuation of low-level and high-level errors is based on different information, it is necessary to differentiate two types of feedback. On the one hand, there is feedback about a specific sensory pattern that signals the success or failure of a movement after termination of that movement. For instance, the distance between the end position of a cursor and a target is visually obtained in a manual aiming task. On the other hand, incoming feedback during the movement can include information about movement execution without revealing clear information about terminal movement success. This information, however, can be used to estimate the terminal movement outcome. Thus, high-level errors in motor tasks can either be evaluated by prediction on the basis of a preceding low-level error transmitted by movement feedback or evaluated by postdiction (detected) after the presentation of feedback about the terminal action effect (success or failure).

To date, and as stated above, ERP components correlating with error processing have been defined and distinguished based on whether internal (movement related) or external (outcome-feedback related) error information is processed. External information has been equated with feedback in general, without differentiating between the content of feedback information. We believe that this distinction is inadequate for research on error processing in motor learning, since the essential drive for learning is information about the success of a movement (with less significance on the locus of information). Hence, we suggest to structure the definition of errors with respect to their information content (high-level or action-effect error vs. low-level or movement error). In addition, ERP components representing the processing of action effect errors need to be differentiated with respect to the type of error valuation: whether the error is predicted or postdicted. The ERP representing prediction could function as a measure of learning. In this regard, we have searched through recent publications to find tasks that examined ERPs of movement error valuation and action effect error valuation in motor tasks. We will first briefly describe the empirical procedures and findings of the studies, which will be followed by our argumentation in favor of an adapted terminology regarding the Ne/ERN and the FRN, concluding with a classification of the selected studies. It is important to emphasize that this review addresses terminological issues related to the examination of the Ne/ERN and the FRN in motor tasks, and it is not its goal to summarize the theoretical and neural underpinnings of these components. Hence, the studies reviewed instead represent examples, and there is no claim made for completeness.

When considering these studies on error valuation in complex motor tasks, it becomes apparent that not all authors adopt the ERP terminology that differentiates between external and internal error information. Hence, we propose an alternative distinction in terminology. First, we suggest to conceptually discriminate between movement descriptions and descriptions of action effects. To keep these characteristics properly separated, we will use Greek letters when referring to movement parameters, and Latin letters when referring to parameters describing action effects. Both are coupled by a (potentially) non-trivial function, e.g. resulting in a denotation like: E =fα1,α2,…,αn, $f\left(\alpha_{1}, \alpha_{2}, \ldots, \alpha_{n}\right),$where E represents the action effect, and f is the model of the true physical laws that connect the movement parameters α1,α2,…,αn $\left(\alpha_{1}, \alpha_{2}, \ldots, \alpha_{n}\right)$to the effect. As stated earlier, an error is typically conceptualized as difference between an intended state and an actual state. Accordingly, when looking at errors in the action effect, it has to be considered what was intended to happen (denoted by E∗), and what actually happened (E ), and, respectively, what is expected or predicted to happen (E^). $(\widehat{E}).$Both, E∗ and are projections into the future (t + Δt) on the basis of information at an earlier point in time (t). For the predicted effect, this can, for example, be denoted E^(t+Δt)=f^α1(t),α2(t),…,αn(t), $\hat{E}(t+\Delta t)=\hat{f}\left(\alpha_{1}(t), \alpha_{2}(t), \ldots, \alpha_{n}(t)\right),$where f is the estimated model of the true physical model (prediction function) that predicts .

After having presented these definitions, we are in a position to define what we think is being reflected in the electrophysiological potentials being measured. According to the theory of hierarchical error processing based on Krigolson and Holroyd, the Ne/ERN and the FRN both reflect processing of high-level errors, while the Ne/ERN is suggested to represent pre-dictive error processing, and the FRN is suggested to correlate with post-dictive error processing. Thus, the FRN should indicate the result of the processing of information about the action effect after it has occurred (note that the observed difference in negativity should not be interpreted as a correlate of the processing of error information per se. Error information is processed in any case. However, negativity solely differs dependent on the result of this processing, i.e. whether or not an error is detected). To be precise, this information should be called terminal action effect feedback (TAEF), so as to differentiate it from the action effect that does not indicate the actual action goal (e.g., ball flight towards the basket in a basketball shot; see Fig.2). The modality in which the information about TAE is transferred, which parameters are actually used to describe the effect, or the source [internal vs. external] is not at all relevant for the question of whether this information can be called TAEF (and the corresponding time locked potentials (TAE)FRN accordingly). Following this definition, the FRN, representing a high-level error E∗ ≠ E, can only be observed after the terminal action effect E has occurred.

Figure 2

The Ne/ERN, instead, reflects a prediction of the terminal action effect. The characteristics of the underlying information for this prediction are again (as in the TAEF information) not relevant in our opinion, as long as the information is not a TAEF. This can be better explained when examining the possible information sources for the prediction. In cases where the movement is identically described as its effect (e.g., button presses, tracking), information about the movement is the same as information about the TAE. Hence, predictive error processing is based on internal information (efference copy of motor command) alone in these cases. In contrast, when movement and action effect are distinct (e.g., in throwing), the prediction can be based on efferent as well as on sensory movement-related information, as long as it is not based on direct information about the action outcome. Here, low-level errors become important. Low-level errors are defined as discrepancies between the intended and the actual movement parameters (Krigolson and Holroyd, 2007b). They can, hence, be denoted as

Low-level errors, those that can be corrected, and as a consequence, do not violate the action goal, have been associated with posteriorly distributed ERP components (N100 and P300; Krigolson and Holroyd, 2007a), but not with fronto-central activity. This is not surprising, since it can be derived from the above equation that low-level errors do not directly project onto the action effect E. To link movement-related information to high-level errors, the prediction function f is required, allowing a projection along two dimensions, i.e. from one level of description to another (from movement to effect), and from the moment (t) in time the data is available to a moment in the future (t + Δt) when the effect will actually occur in reality. Hence, a high-level error occurs when fα1(t),α2(t),…,αn(t) $f\left(\alpha_{1}(t), \alpha_{2}(t), \ldots, \alpha_{n}(t)\right)$inevitably leads to an E (t + Δt) that is different from E∗(t + Δt). However, since E (t + Δt) is not known at time t, the only reasonable way to compare the actual with the intended effect is to replace E (t + Δt) by an estimate, i.e. E^(t+Δt),$\hat{E}(t+\Delta t),$which can be considered a prediction of the TAE. In cases where the predicted TAE is evaluated as an error (high-level error), fronto-central activity in form of an Ne/ERN should be observable. Thus, to clarify once more: The labelling of error-related brain activation is independent from the modality and the parameters used; the defining distinction is whether this activation is based on a predicted effect or not.

In summary, errors can occur at two different levels: the movement execution level and the action effect level. Depending on the task, these levels can coincide or be viewed separately. Figure 2 illustrates different variables of the error valuation process in two different tasks, basketball shooting and button pressing. The main difference between the tasks is the separation between action and movement. While movement and action are identical in button pressing, the shooting action is not terminated with the end of the shooting movement represented by ball release. The ball still needs some time (about 1 second) to travel towards the basket after it has been released from the shooter’s hand. Hence, TAEF is delayed with respect to the end of the movement. Prediction of the TAE can take place at roughly three different time points or intervals. First, internal information from the efference copy of the motor command is available after termination of movement planning for a comparison of the intended effect E∗(t + Δt) with the predicted effect E^(t+Δt).$\hat{E}(t+\Delta t).$Second, sensory information about the movement is collected during movement execution. The matching of intended α1∗(t),α2∗(t),…,α∗n(t) $\left(\alpha_{1}^{*}(t), \alpha_{2}^{*}(t), \ldots, \alpha^{*}{ }_{n}(t)\right)$with actual movement parameters α1(t),α2(t),…,αn(t) $\left(\alpha_{1}(t), \alpha_{2}(t), \ldots, \alpha_{n}(t)\right)$gives rise to an updated prediction of (t + Δt) and, hence, an updated comparison with E∗(t + Δt).

It is important to recall that a high-level error is only detected in the case that a movement error cannot be corrected anymore (which is the case after the moment of ball release in throwing). A third comparison of E∗(t+Δt) withE^(t+Δt) $E^{*}(t+\Delta t)\,\text{ with} \,\widehat{E}(t+\Delta t)$can occur during the ball flight towards the basketball rim. Although certainty of the prediction increases with movement termination and the closer the ball gets to the rim, error valuation will remain a prediction as long as TAEF has not yet started (i.e., the ball has not yet passed the rim level). In contrast, TAE in the button press task can only be predicted after movement planning has been terminated. Deviations from the desired movement parameters that occur during movement execution cannot be processed and used to derive a predictive valuation of the TAE before the action is actually terminated, due to the shortness of the movement. Since movement end and action end are identical, error valuation after having pressed the button falls into the postdiction interval.

Depending on which error level or type (action effect or movement execution) one is interested in, trials have to be categorized into error trials and successful trials on their respective levels. That is, when analyzing movement errors, their categorization must not be based on the outcome of the effect level, but rather according to the average deviation of the movement trajectory or a similar characteristic. Thus, when an ERP signal emerges during the prediction interval (i.e., before TAEF) and correlates with an error on the effect level, it predicts this TAE-error. In contrast, when an ERP signal emerges during the prediction interval and correlates with a movement error (e.g., a trajectory error), it does not necessarily have a predictive function, since movement errors do not have to end in TAE-errors. This is confirmed by the study by Krigolson and Holroyd (2007a), who found a Ne/ERN that correlated with effect errors in a manual aiming task when the aiming error could not be corrected. In contrast, when participants were able to correct their aiming movements, the Ne/ERN was absent, and ERP signals with more parietal distributions were observed (see also Krigolson and Holroyd, 2006). In this regard, we propose the use of the term error-related for neural correlates of error valuation processes predicting a TAE-error, independent of the input source, the exact timing in the interval, or the task. Conversely, neural correlates would be termed feedback-related if they emerge after the TAEF in the postdiction interval.

To conclude, the classical terminology that differentiates the Ne/ERN and the FRN with respect to internal and external error information is unambiguous in tasks where low-level movement errors and high-level action effect errors coincide. In cases where a movement error leads to an action effect error only after a delay (e.g., tracking, throwing, shooting), ERPs can indicate the movement error, the prediction of the action effect error, and/or the detection (postdiction) of the action effect error. Similarly, error information can vary between efferent (internal), efferent and afferent (external), as well as solely afferent sources. Thus, it is important that these characteristics of error production and valuation are precisely defined in empirical studies, and that terminology is consistently used. Our proposal to choose Ne/ERN for the prediction of an action effect error and FRN for the postdiction of an action effect error is an attempt to clarify ERP terminology, especially for tasks where the action effect is dependent on the movement execution. It has to be pointed out that it was not within the scope of this focused review to address movement errors that do not lead to action effect errors (as investigated in de Bruijn et al., 2003; Krigolson and Holroyd, 2006, 2007a, and only observed in Krigolson and Holroyd, 2007b).

An alternative use of terminology to that presented here is that the Ne/ERN reflects internal sensory-prediction errors (which are action/execution oriented; motor control models), when the FRN reflects reward prediction errors (which are output-oriented regarding reinforcement learning theory; decision making theories; Torrecillos, et al. 2014). Hence, if a detected movement error (on the basis of a sensory prediction error signal) predicts a failure to reach an action goal (negative reward prediction), all ERP signals representing predictions regarding action outcomes/rewards should be termed FRN. This, in turn, would limit the Ne/ERN to movement errors with no clear predictive function regarding outcomes. This would only be the case in the studies of de Bruijn and colleagues (2003), and Krigolson and Holroyd (2007b) who, however, did not find negative fronto-central ERP signals that correlated with movement errors. Krigolson and Holryod (2006) also reported movement errors that did not lead to action outcome errors, because they could be corrected. However, they did not find fronto-central signals either, but negative occipital-parietal ERPs. Thus, in line with a postulation by Cavanagh (Cavanagh et al., 2012) that the Ne/ERN, the FRN and also the N2 reflect a single frontal midline theta-rhythm sensitive to mismatch signal in the service of behavioral adaptation, one could argue to confine the terminology of the error-related ERP signals to either the Ne/ERN or the FRN. However, it is unquestionable that temporally variable error-related ERP signals exist (early after movement onset vs. following outcome feedback), which also differ with respect to expectancy/surprise and error size, as well as within the learning process. Taking these differences into account, it seems advisable to express them through distinct terminology.