The Neural Correlates of Motor Skill Automaticity

Introduction

A universal feature of human behavior is that well practiced tasks can be performed with relatively little effort or cognitive control, whereas novice performance of the same task may require intense and effortful cognitive control (James, 1890). The ability to perform skilled tasks without the need for executive control is referred to as “automaticity,” and it has been extensively investigated by cognitive scientists over the past 40 years. A defining characteristic of automaticity is the ability to perform the automatized task with little or no interference by a demanding secondary task (Posner and Snyder, 1975; Logan, 1979). The neural basis of automaticity is of great interest, but little is currently known about how neural activity changes as a task becomes automatized (but see Wu et al., 2004).

In the present study, we trained subjects extensively on an implicit sequential reaction time (SRT) task that has been used both behaviorally and with neuroimaging to examine motor skill learning. In this paradigm (Nissen and Bullemer, 1987), subjects are presented with a four-alternative spatial choice reaction time task; unbeknownst to them, on some blocks of trials, the locations occur in a particular sequence, whereas on other blocks, the stimuli are presented in a pseudorandom order. Subjects will generally respond more quickly on all blocks as training progresses. In addition, there is a greater decrease in response time (RT) for sequence blocks compared with pseudorandom blocks, even if subjects cannot explicitly report the sequence (Willingham et al., 1989) (but see Perruchet and Amorim, 1992). We examined performance of the SRT task under both single-task and dual-task conditions. Subjects were scanned using functional magnetic resonance imaging (fMRI) on two occasions, with 3 h of SRT training in between these scanning sessions. Dual-task probe blocks were presented during both training and imaging, and degree of automaticity was defined as the cost of performing the secondary task.

The lateral prefrontal cortex has been particularly implicated in the support of cognitive control operations (Norman and Shallice, 1986; Miller and Cohen, 2001). In particular, there is evidence that lateral prefrontal regions are recruited under conditions of dual-task performance (Dreher and Grafman, 2003) and in situations requiring task switching (Sohn et al., 2000; Wylie et al., 2003). These results suggest that the development of automaticity should be associated with decreasing activity in lateral prefrontal regions. At the same time, previous studies of both motor (Grafton et al., 1994; Van Horn et al., 1998; Penhune and Doyon, 2002) and perceptual (Poldrack et al., 1998; Poldrack and Gabrieli, 2001) skill acquisition have found learning-related increases in activation in both cortical and subcortical regions, including the neostriatum (putamen for motor skill learning and caudate for perceptual skill learning). We tested for regions of either increased or decreased activation with automaticity to test the hypothesis that different regions may exhibit divergent changes related to the decreased need for cognitive control with practice.

Materials and Methods

Subjects. Fourteen normal young adult subjects participated in this study for payment (mean age, 20.7; eight females). All subjects provided informed consent according to procedures approved by the University of California, Los Angeles Human Subjects Committee. One additional subject was excluded because of failure to complete the second imaging session.

Experimental design. Each subject participated in two scanning sessions, separated by 1 or 2 d. During the intervening period, subjects participated in three training sessions. Figure 1 presents a schematic of the task design.

The task was an SRT (Nissen and Bullemer, 1987), in which subjects were presented with a visual target in one of four screen locations and asked to respond as quickly and accurately as possible to this location by pressing one of four spatially compatible buttons. The visual target (“X”) was presented for 1 s, with a 1.25 s interstimulus interval. During some blocks (“sequence” blocks), the stimulus locations followed a consistent 12-item-long sequence (unbeknownst to the subject). This was a second-order conditional sequence, meaning that every individual location and every first-order transition appeared equally often. During other blocks (known as “pseudorandom” blocks), items were presented in a pseudorandom order, with the constraint that all individual targets appear equally often as in the sequence blocks (Reed and Johnson, 1994). This design ensures that any differences in behavior or activity between sequence and random blocks must reflect the learning of higher-order sequential knowledge rather than lower-order statistics.

Subjects performed the SRT task under both single-task and dual-task conditions. Under the single-task condition, only the SRT was performed. Under dual-task conditions, the SRT was performed concurrently with a tone-counting task. A single tone was presented each time a visual target appeared; across subjects, 50-70% of the tones were high pitched (1000 Hz), with the remaining tones at 500 Hz. Subjects were asked to count the number of high tones occurring during each dual-task block; during scanning runs, they did not report these counts, whereas during training runs, they reported them by entering the number using a keyboard.

Each scanning session consisted of two functional imaging runs. During each of these runs, there were six blocks (lasting 45 s each), during which the subject performed the SRT task, separated by blocks of rest lasting 15 s each; each run also began with 15 s of rest. Each scan included the following: two single-task sequence blocks, two single-task pseudorandom blocks, one dual-task sequence block, and one dual-task pseudorandom block (Fig. 1). Two orders were used, with a different order for each of the two runs in each session (with order counterbalanced across runs and across sessions). Key-press responses and response times were collected using a custom four-button box.

During the intervening period between scanning sessions, subjects participated in three training sessions. During each session, they performed 12 blocks of the SRT task, each of which included 60 sequence trials and 60 pseudorandom trials. All blocks were performed under single-task conditions, except for blocks 3, 4, 9, and 10, which were performed under dual-task conditions as described above (Fig. 1).

Imaging procedures. Imaging was performed using a 3.0 T GE MR scanner with Advanced NMR gradients (GE Medical Systems, Milwaukee, WI). Blood oxygenation level-dependent-sensitive functional images were collected using a gradient-echo echo-planar pulse sequence [repetition time (TR), 3000 ms; echo time (TE), 45 ms; 64 × 64 matrix; 19 slices, 4 mm thick, 1 mm interslice gap]. Two runs of fMRI acquisition were performed for each subject during each scanning session, lasting 363 s each (encompassing 360 s for the task paradigm and an additional 3 s at the end of the run), with an additional four images at the beginning of the run discarded to allow T1 equilibration. Before functional scanning, we collected a coplanar anatomical image of the whole brain using a T2-weighted spin-echo echo-planar sequence to aid intersubject registration (TR, 4 s; TE, 54 ms; 128 matrix; 26 slices, 4 mm thick, 1 mm interslice gap).

fMRI data analysis. Preprocessing and statistical analysis of the data were performed using SPM99 software (Wellcome Department of Cognitive Neurology, London, UK). The data were first analyzed for outlier images (caused by scanner spikes); such images were replaced with the mean of the two temporally adjacent images in the time series, with a total of 12 images replaced across the entire study. Preprocessing with SPM99 included motion correction and spatial normalization to a custom template approximating the MNI 305 stereotactic space (Montreal Neurological Institute), using linear affine registration followed by nonlinear registration using cosine basis functions, resampling to 3 mm cubic voxels. Normalization was performed by aligning the coplanar anatomical image to the template and then applying the same normalization parameters to the aligned functional images. Statistical analysis was first performed for each individual subject using a general linear model in SPM99. The design was modeled as a boxcar convolved with a canonical hemodynamic response, with each block modeled separately. Low-frequency signal components (120 s cutoff) were treated as confounding covariates. The model fit was performed individually for each subject, and contrast images were generated for each of the four event types (against the explicitly unmodeled baseline/fixation). The contrast images were spatially smoothed with an 8 mm Gaussian kernel and then used in a second-level analysis treating subject as a random effect. The second-level statistical maps were thresholded at p = 0.005 (uncorrected); control for familywise error was achieved using the cluster-level Gaussian random field results, resulting in a volumewise corrected error rate of p < 0.05.

Region-of-interest (ROI) analyses were performed using the SPM ROI Toolbox (http://spm-toolbox.sourceforge.net). Regions were defined functionally based on voxelwise statistical maps, by growing a sphere (6 mm radius) around the local maxima in each cluster. For clusters of additional interest, the effect of each condition versus baseline was estimated for each subject by modeling the data using a finite impulse response model and averaging the response over the length of the trial block. This provided an estimate of the average signal change induced by each condition, without any assumptions regarding the shape of the response.

Results

Discussion

The present study examined changes in neural activity associated with the acquisition of automaticity in a motor learning task. There were three major findings. First, novice performance of an SRT task with a concurrent working memory load resulted in activation of a network, including PMv/IFG and DLPFC and right caudate body, compared with performance without a concurrent task. This suggests that these regions are critical for cognitive control during early learning. Other regions, including dorsal premotor cortex and SMA, were equally active regardless of the presence of the secondary task, consistent with a more fundamental role in motor programming. Second, practice was associated with the development of automaticity (as indexed by the elimination of dual-task interference), and this change was associated with decreased activation in PMv/IFG, DLPFC, and right caudate regions. Finally, basal ganglia regions [including putamen, globus pallidus (GP), and caudate head] exhibited an interaction of sequence type and time, such that training resulted in decreased activity for sequential trials but not for pseudorandom trials. These results provide direct evidence that these regions are involved in the acquisition of motor sequence knowledge, consistent with previous imaging (Grafton et al., 1995; Rauch et al., 1997) and neuropsychological (Knopman and Nissen, 1991; Willingham and Koroshetz, 1993) results.

It is of note that the main findings regarding automaticity in this study were general to both sequence and random conditions in the SRT task. This is supported by the lack of an interaction between dual-task effects and sequence versus random conditions on response time during training and by the lack of differences in the neural signature of dual-task effects for sequence versus random conditions. These results diverge from previous behavioral studies, which found that dual-task conditions resulted in impaired acquisition (Curran and Keele, 1993) and/or expression (Frensch et al., 1998) of sequence knowledge. It is possible that this difference from previous results arose from a procedural difference between the present study and those previous studies. Nearly all previous studies presented the tone at randomly spaced times after presentation of the SRT stimulus and approximately coincident with the motor response or in the interval between the motor response and the next stimulus (cf. Hsiao and Reber, 2001), whereas we presented the tone concurrently with the onset of the SRT stimulus. Hsiao and Reber (2001) systematically varied the occurrence of the tone within the response-stimulus interval and found the greatest interference late in this interval. This suggests that a consistently timed tone before the response may have reduced the disruption of sequence timing. Instead, the presence of the tone appeared to exert a general effect on motor responding during earlier sessions.

Several previous studies have examined neural changes associated with extensive (multisession) training on SRT tasks. Karni et al. (1995) trained subjects on an explicit five-element finger movement sequence for 10-20 min of practice per day for 5 weeks. No secondary task was performed, but the fact that subjects continued to practice for 2 weeks beyond asymptote suggests that automaticity was likely achieved. This study used a surface coil and only imaged a small region of motor cortex, in which an increase in activity related to sequence learning was observed compared with an untrained motor sequence. No increase was observed in motor cortex in the present study, but this may have been attributable to the degree of smoothing applied to the data, which may have obscured small, localized changes in activity. Wu et al. (2004) scanned before and after an explicit sequence learning paradigm presenting two sequences of different lengths to each subject (4 or 12 elements) and administering a secondary task to index the degree of automaticity. Automaticity was achieved within 2-4 h of training on both sequences. A number of regions showed decreased activation after training, including bilateral prefrontal and parietal regions, cerebellum, and caudate nucleus; as in the present study, no increases in activation were observed after training. However, this study did not directly image SRT performance under dual-task conditions. Thus, our results extend previous findings by demonstrating that automaticity is associated with specific decreases in activation in a fronto-striatal network during dual-task performance.

Although the focus of our study was on motor skill learning in general, there were some modest differences in activity between sequence and pseudorandom conditions. Several previous studies have found patterns of sequence-specific activation in the SRT that differ depending on whether sequence learning was implicit or explicit (Grafton et al., 1995; Rauch et al., 1995; Hazeltine et al., 1997; Honda et al., 1998). In the present study, awareness of the repeated sequence was not assessed. The second-order conditional sequence used in the present study and the mixing of sequence and random blocks was likely to have reduced sequence awareness, although it is possible that some subjects acquired at least partial explicit sequence knowledge. Our neuroimaging results overlap to some extent with previous studies of implicit sequence learning, in that there was a strong trend for sequence-specific caudate activation early in training. It is unclear whether the possibility of explicit sequence knowledge in our subjects had a major impact on the pattern of results, given more recent findings of Willingham et al. (2002) that regions active during implicit learning were also active during explicit learning (but not vice versa).

Because the development of automaticity in this task was not sequence specific, it must reflect changes in the processing of more basic invariant features of the tasks that may be shared by other motor skill tasks. One major aspect of task-general learning in our paradigm may be the tuning of spatially compatible stimulus-response mappings. In this context, it is interesting that previous studies have associated ventrolateral prefrontal cortex with the learning, maintenance, and controlled retrieval of stimulus-response associations (Asaad et al., 1998; Passingham et al., 2000; Bunge et al., 2003). The increased activity of bilateral ventrolateral prefrontal cortex (PMv/IFG) in our dual-task condition before training suggests that dual-task conditions may require greater control over stimulus-response mapping rules, even in cases in which the mappings are spatially compatible. With extended practice, these mappings appear to be retrieved automatically and thus independently of ventrolateral prefrontal circuitry.