Transcranial Direct Current Stimulation enhances Motor Skill Learning but not Generalization in chronic Stroke

Abstract

Background

Motor training alone or combined with transcranial direct current stimulation (tDCS) positioned over the motor cortex (M1) improves motor function in chronic stroke. Currently, understanding of how tDCS influences the process of motor skill learning after stroke is lacking.

Objective

To assess the effects of tDCS on the stages of motor skill learning and on generalization to untrained motor function.

Methods

In this randomized, sham-controlled, blinded study of fifty-six mildly impaired chronic stroke patients, tDCS (anode over the ipsilesional M1 and cathode on the contralesional forehead) was applied during five days of training on an unfamiliar, challenging fine motor skill task (sequential visual isometric pinch force task). We assessed online and offline learning during the training period, and retention over the following four months. We additionally assessed the generalization to untrained tasks.

Results

With training alone (sham tDCS group), patients acquired a novel motor skill. This skill improved online, remained stable during the offline periods and was largely retained at follow-up. When tDCS was added to training (real tDCS group), motor skill significantly increased relative to sham, mostly in the online stage. Long-term retention was not affected by tDCS. Training effects generalized to untrained tasks, but those performance gains were not enhanced further by tDCS.

Conclusions

Training of an unfamiliar skill task represents a strategy to improve fine motor function in chronic stroke. tDCS augments motor skill learning, but its additive effect is restricted to the trained skill.

Introduction

Worldwide, stroke is the leading cause of persistent motor disability. Despite significant functional recovery in the first three to six months after injury^1,2 many patients remain substantially limited by chronic motor impairment that interferes with independent living ^3,4, prompting a need to identify therapeutic interventions that could promote recovery well into the chronic stage.

The central tenet in stroke rehabilitation is motor training ^5,6. In healthy subjects, training leads to motor skill changes that take place during practice (online learning), in between training sessions (offline learning) and that may be maintained over time (long-term retention) ^7–10. In patients with chronic stroke, training also improves motor skills ^11–18. However, the integrity of the different learning stages over multiple training sessions has not been explored in detail after stroke.

Transcranial direct current simulation (tDCS) modulates cortical excitability ^19,20 and promotes synaptic plasticity in a polarity-dependent fashion ^21,22. In healthy humans, tDCS with the anode placed over the primary motor cortex improves motor skill learning ^9,10,23,24. tDCS combined with physical or occupational therapy can also improve functional motor outcomes after stroke ^25–28. However, clear recommendations for therapeutic use of tDCS are currently not given due to heterogeneity in outcome variables, interventions and study populations ²⁹. Moreover, these previous studies did not specifically evaluate training or tDCS effects on the stages of motor skill learning during therapy, or their interactions. After stroke, tDCS with the anode placed over the ipsilesional primary motor cortex also enhances skill learning within a single session ^30,31 with post-training skill levels maintained for about one week. However, replicability of this finding has been questioned ³². The influence of tDCS on stages of motor skill learning over multiple days (online and offline learning and long term retention) are unknown. This information is needed to gain insight into the mechanisms by which noninvasive brain stimulation influences motor learning after stroke and to optimize dosing and timing of interventional brain stimulation protocols in neurorehabilitation.

Here, we investigated the influence of repeated training, tDCS and their interaction on stages of multi-session motor skill learning after chronic stroke. Our primary hypothesis was that tDCS would enhance total learning. Additionally, we explored whether skill improvements generalize to untrained tasks compared to a no-training/no-tDCS group, and whether generalization is affected by tDCS. Finally, we assessed patient characteristics that may affect total learning or the response to tDCS, such as lesion anatomy or the BDNF val66met polymorphism.

Methods

Patients and Experimental Design

This prospective, randomized, sham-controlled, blinded study was approved by the Institutional Review Board at the NINDS, National Institutes of Health (NIH) and the local Ethics Committee of the University of Freiburg. The study was preregistered (www.clinicaltrials.gov, NCT00314769) and conducted in accordance with the declaration of Helsinki. All patients gave written informed consent. Eight patients were tested at the NIH site; forty-eight patients were subsequently tested in Freiburg, Germany using the same experimental set up.

Clinical assessment

Chronic stroke patients were screened by a detailed clinical interview and exam, standardized questionnaires and genotyping for the BDNF val66met polymorphism (DNA analysis described previously²¹). Details are shown in Figure 1B, Table 1 and in Supplementary methods. Inclusion criteria were: 1) age 18–80 years, 2) unilateral, first ever ischemic stroke more than 3 months before study enrollment, 3) mild to moderate hemiparesis with residual hand function sufficient for task performance, 4) clear hand preference as assessed by the Edinburgh Handedness Inventory³³ and 5) sufficient cognitive function to comply with study requirements. Patients at the Freiburg site received a structural MRI to characterize lesion anatomy. NIH patients did not receive scans due to contraindications, personal reasons or logistic reasons.

Figure 1. Study design.

A, Patients trained for five consecutive days and were then subjected to 5 follow-up visits. During training they received real (40µA/cm2) or sham tDCS for 20 minutes per day with the anode targeting the primary motor cortex (M1) of the affected hemisphere. A training session consisted of 5 blocks of 20 trials of the sequential visual isometric pinch force task (SVIPT). At the beginning of day 1, the end of day 5 and all follow-ups the Jebsen Taylor hand function test (JTT) and the Grooved Pegboard test (GPT) were also performed. The no-training/no-tDCS group performed only the generalization tasks without receiving training or tDCS. B, CONSORT flow diagram of the study. C, The summarized binary lesion overlay for each study group is shown, blue=sham tDCS, red=real tDCS, green= no-training/no-tDCS group. Color key (dark to bright) ranges from 1 to 6 patients. The majority of patients had an infarction mainly located along the pyramidal tract and particularly in the internal capsule. Stimulation groups were well matched for their lesion anatomy and volume. VLSM revealed no group differences.

Table 1.

Patient information, psychophysical and baseline behavioral data

	SHAM	REAL	NO TRAINING/ NO tDCS	p (group)
Patient information
N	18	18	14
Age (years)	61.6±3	61.9±3	64.7±2	0.707
Gender (Male/Female)	67/33%	83/17%	57/43%	0.257
Time after stroke in months	43.7±12	47.9±19	22.9±4	0.449
Affected Hand R/L	44/56%	50/50%	50/50%	0.931
Concordance (hand affected dominant/non-dominant)	50/50%	61/39%	64/36%	0.680
Edinburgh Handedness Score	83.8 ± 10	77.9 ± 11	90.0 ± 9	0.739
UEFM (/66 maximum)	59.3 ± 1	58.7 ± 1	58.9 ± 1	0.951
Modified Ashworth Scale	0.54 ± 0.2	0.69 ± 0.2	0.43 ± 0.1	0.502
Sensory Touch Test (/50 maximum)	43.8 ± 2	37.0 ± 3	43.0 ± 2	0.146
MMSE (/30 maximum)	29.1 ± 1	28.9 ± 0.3	28.4 ± 0.4	0.484
BDI-II Day 1 (/max. 63)	10.7 ± 2.0	8.7 ± 1.8	7.4 ± 1.4	0.548
Met carriers BDNF val66met polymorphism	39%	33%	36%	0.911
Stroke volume (ml)	15.2 ± 4	14.4 ± 7	18.5 ± 7	0.880
Pure subcortical stroke	50%	50%	50%	n/a
Baseline behavioral data
Skill measure day 1	−2.1 ± 0.3	−2.2 ± 0.2	n/a	0.752
JTT time (seconds)	53 ± 4	55 ± 5	62 ± 9	0.583
GPT time (seconds)	326 ± 69	298 ± 62	266 ± 58	0.815
GPT error count (no.)	2.7 ± 0.7	2.8 ± 0.6	2.1 ± 0.6	0.718
Psychophysical data
tDCS condition guess (felt stimulated)	80%	70%	n/a	0.729
Side effects of tDCS*	33%	11%	n/a	0.109
Average researcher’s influence (VAS, −5 to +5)	0.6 ± 0.9	0.8 ± 0.4	0.6 ± 0.3	0.569
Average alertness before training (VAS, 1–10)	3.5 ±0.7	3.4 ± 0.4	4.4 ±0.5	0.123
Average sleep duration (hrs) before training	7.4 ± 0.5	6.6 ± 0.3	7.1 ± 0.5	0.086
Average time of testing (/24hrs)	11.1 ± 0.5	11.5 ± 0.5	11.3 ± 0.75	0.858

Group means and SEM or percentages per group are given.

UEFM: upper extremity Fugl-Meyer score, MMSE: Mini mental state examination, BDI-II: Beck Depression inventory-II. VAS: visual analogue scale.

^*Mild side effects occurred in eight patients: headache (n=3, sham), migraine (n=1, real tDCS), phosphene (n=1, sham), abdominal pain (n=1, sham), retching (n=1, sham), and tingling sensation of the unaffected hand (n=1, real tDCS).

Study Design

The study consisted of five days of consecutive training and five follow-ups over four months (Fig. 1A). Sessions took place between 8 AM and 3 PM in the same environment and instructed by the same investigator (M.H., J.R., H.M.S.), and time of the session was held constant for each patient (±1 hours allowed). All patients were naïve to the motor tasks utilized and had not taken part in tDCS experiments before. Patients were randomized into the two training groups (real tDCS and sham tDCS). Allocation to a tDCS condition followed a balanced randomization list prepared prior to the experiment. A third group (no-training/no-tDCS) was added during the course of the experiment, with adaptive randomization for the allocation into training or no-training condition (Figure 1B, supplementary methods).

Real and sham tDCS groups (n=18 each): Patients practiced a modified version of the sequential visual isometric pinch force task (SVIPT¹⁰, Fig. 1B) with their paretic hand for approximately 45 min per day (5 × 20 trials with short breaks between blocks to avoid fatigue). As in a previous study¹⁰, patients squeezed a force transducer between thumb and index finger to move a cursor as quickly and accurately as possible through a fixed pattern of five target gates on a computer screen. Patients received visual feedback based on the cursor location, which partially allowed them to interpret their success, e.g. in case a target was not hit correctly (over-/undershoots). There was no additional feedback provided by the investigators. After the first training block of each day, real or sham tDCS was applied for 20 min to the M1 of the affected hemisphere. Stimulation started after block 1 to keep the first training block free of online stimulation effects for the analysis of offline learning. Training outlasted the stimulation on the first day, but as patients became faster over time, training blocks were spaced to align with stimulation duration. To assess the generalization of hand motor skill learning to upper extremity function, patients performed the Jebsen Taylor hand function test (JTT) and the Grooved Pegboard Test (GPT, Fig. 1B) at the beginning of the first and the end of the fifth training day, and after SVIPT assessment at each follow-up. The JTT was performed three times and the GPT was performed twice. These tests were chosen because they are both executed unimanually, require pinch and grip movement elements (the pinch showing similarity to the SVIPT) and both have been reliably used in stroke patients ^34–37.

No-training/no-tDCS group (n=14): Patients performed only the JTT and GPT on days 1, 5, and at each follow-up time point. They did not undergo SVIPT training or tDCS. This group was implemented to discern repetition effects of the JTT and GPT on their performance from generalization effects of SVIPT training and tDCS.

Transcranial Direct Current Stimulation

We followed our published standard operating protocol ³⁸ for the determination of the cortical hotspot using single-pulse transcranial magnetic stimulation (here: the M1 hotspot at which consistent MEPs were evoked in the contralateral (paretic) first dorsal interosseus muscle). MEPs were present in all patients. The anode (25 cm²) was centered over the ipsilesional M1 hotspot while the cathode (25 cm²) was placed over the contralateral supraorbital area. In the real tDCS condition, 20 minutes of 1 mA current was delivered (current density 0.4 A/m², total charge 0.048 C/cm²). Sham tDCS was ramped up and then down over 30s to induce similar scalp sensations as real stimulation. Stimulation was delivered in a single-blind manner (Phoresor II, Iomed) at the NIH site and double-blinded (Neuroconn DC-Stimulator plus, with study mode) at the Freiburg site. tDCS blinding success was assessed by asking patients to guess whether they received active or inactive stimulation.

Refinement of the mathematical skill model

For determination of motor skill, the speed-accuracy assessment described previously¹⁰ was refined to incorporate more fine-grained detail about force precision. Movement time (time from movement onset until reaching the fifth target) and target error rate (percentage of missed targets per training block) were combined into a single index, the skill parameter. The speed-accuracy trade-off function specific to stroke patients was obtained from an unrelated subset of patients to mathematically model skill (see supplementary data). The function generated was as follows

$$a=\frac{1-\mathit{\text{target error rate}}}{\mathit{\text{target error rate}}\ast (\mathit{\text{ln}}{(\mathit{\text{movement time}})}^{3.43})}$$

where a is the skill parameter incorporating a performance’s target error rate and movement time, and 3.43 is the constant. “Skill” reflects the log-transformed skill parameter (ln(a)).

Analysis of skill learning and generalization

Overall training effect

After comparing baseline skill (block 1, day 1) to ensure comparability of the groups, we assessed the success of training by evaluating time (day 1 block 1, day 5 block 5) and group differences (sham, real tDCS) in skill as well as their interaction using a repeated measures ANOVA.

Total learning (primary outcome)

As before ¹⁰, the primary outcome measure was total learning, i.e. the sum of skill changes occurring by the end of training (last block on day 5). These skill changes were compared between groups.

Learning stages

Online learning was defined as the sum of skill differences between the first and last block of each of the 5 training days. Offline learning was defined as the sum of skill differences between the last and first block of consecutive days. Retention was defined as the delta between the last block on day 5 and the block on day 113, and the rate of forgetting was the time-weighted slope of skill over all follow-ups after training. For all of these measures, positive values correspond to a skill gain, while negative values express a skill loss. To relate the skill remaining at day 113 to the individual skill level achieved by the last block of day 5, retention was additionally expressed as a proportion of skill retained at day 113.

Generalization

For each patient and assessment, we averaged the overall time to complete each JTT repetition, and the time and errors to complete the GPT. For the JTT and GPT, percentage change in speed (total time to complete the task) was assessed for day 5 and each of the follow-ups (day n) as

$$\%\mathit{\text{speed change}}=(1-(\frac{\mathit{\text{day n speed}}}{\mathit{\text{day}}1\mathit{\text{speed}}}\left)\right)\ast 100$$

so that a positive value implies a reduction in time and therefore an improvement in speed. For GPT accuracy, the absolute change (delta) in errors was calculated across time points as

$$\mathit{\text{absolute accuracy change}}=(\mathit{\text{day}}1\mathit{\text{errors}}-\mathit{\text{day n errors}}),$$

so that a positive value implies a reduction in errors and therefore an improvement in accuracy. We assessed GPT errors as absolute change because several patients made zero errors, precluding a relative assessment. These parameter changes were compared between the two stimulation groups and the no-training/no-tDCS group and across time in a multivariate model.

Online learning on the first day

To enable comparison with the previous literature, learning within the first training day (the skill difference between the first and last block of day 1) was analyzed.

Cumulative learning probability

We specifically addressed the cumulative learning probability by Kaplan-Meier analysis. A skill change of 1 was used as “time to event.” This criterion was chosen based on visual inspection of the skill changes (see Fig. 2A) which corresponds to the value, which sham patients clearly exceeded during training and maintained long-term. A skill change of 1 thus represents a minimum state of skilled proficiency which all subjects did not naively have, but could potentially achieve with training (disregarding a possible tDCS-related placebo effect in the sham condition). For each patient, we identified the time point (training block) at which this proficiency was achieved, and the proportion of patients per group was counted and/or censored. Censoring applied to those patients in whom the time to event could not be determined because skill remained below our threshold criterion until the end of training. A binomial Effect Size Display (BESD³⁹) was used to describe by which percentage the likelihood to reach a skill gain >1 increases under the influence of tDCS.

Figure 2. Measures of motor skill learning during training and follow-up period.

A, Skill learning curve over five training days (five blocks per day) and the follow-up period. The patients receiving real tDCS (red dots) showed significantly greater improvement of motor skill and outperform sham stimulated patients (blue dots) at all time points. The majority of acquired skill was retained after the end of training. At day 113, the real tDCS group showed greater remaining skill compared to sham tDCS. B, Higher total learning in the real tDCS group (red) was predominantly due to greater online (within session) learning. C, The proportion of skill retained at day 113 was similar in the two stimulation conditions, suggesting that real tDCS did not per se affect long-term retention. Significance: *p<0.05. All data are shown as group mean ± SEM.

Relation between online and offline learning

We explored the relation between online and offline skill changes and asked whether tDCS would specifically perturb it. Correlations between online and offline skill changes were calculated per group. To discern an effect of tDCS on the online-offline relation, the two correlation coefficients were transformed with the Fisher’s Z-transformation and the Z-scores were compared.

Changes in movement time, target error rate and movement smoothness

To further discern which aspects of skill had changed over training, movement time, target error rate, and movement smoothness were assessed as separate measures. Smoothness characterizes variations in the movement jerk, the third derivative of the recorded movement force profile, which is expressed as the Integrated Squared Jerk (ISJ) ⁴⁰. The normalized jerk (NJ) multiplies the ISJ such that it results in a unitless number (NJ= √((0.5*ISJ*movement time⁵)/(Cursor path length²))); ^40,41. Smoothness is reported as the inverse value (1/norm jerk). We evaluated time and group effects on each of the three parameters at day 1 and 5.

Correlation of total learning with patient characteristics

Baseline performance, time since stroke, stroke lesion volume, lesion location, stroke concordance (dominant hand affected) and the BDNF val66met polymorphism may conceivably influence the patients’ capacity to train, learn, or respond to tDCS ^1,21,32,42. We correlated the first three parameters with total learning to address the role of these potential confounders. For lesion location, concordance and the BDNF val66met polymorphism, total learning was compared between subgroup of patients. Lesion location on the MRI was rated as either subcortical or as combined lesion (subcortical plus cortical). The BDNF val66met polymorphism state was categorized as either BDNF val/val or as BDNF met carrier (val/met or met/met).

MRI stroke lesion assessment

Lesion detection, lesion overlays (Fig. 2) and voxel-based lesion symptom mapping regarding the lesion anatomy were performed as described previously⁴³ (see Supplementary Methods).

Psychophysical assessment

Side effects, sleep duration, researcher’s influence (visual analogue scale (VAS) from −5 (very negative) to +5 (very positive)), and alertness (VAS from 1 to 10 (“perfectly alert” to “extremely tired”) were recorded daily. To detect any longer lasting mood changes, patients completed the Beck Depression Inventory (BDI) and the Positive and negative affect scale (PANAS) on day 1, 5, and 8.

Statistical Analysis

Sample size calculation is provided in supplementary methods. Following the Shapiro-Wilk test for normal distribution group differences were assessed by two-tailed t-test or ANOVA comparisons with Bonferroni-Holm corrected post-hoc tests. Two-sided significance level was set to p< 0.05. Effect size is given as Cohen’s d (<0.5 indicating a small, 0.5–0.7 a moderate and >0.7 a strong effect)⁴⁴.

The overall training effect was assessed using a repeated measure ANOVA with factors TIME (d1 block 1, d5 block 5) and GROUP (sham, real tDCS). The primary outcome parameter was total learning, which was assessed as group difference between real and sham tDCS group (t-test). Further t-tests were used to assess group differences in baseline skill, online learning, offline learning, retention measures, learning on the first day. JTT speed, GPT speed and GPT accuracy changes were compared by two multivariate ANOVAs: First, we addressed time and overall group differences related to training state (training versus no training) as well as their interactions. Second, time and overall group effects related to the tDCS condition (sham versus real tDCS) as well as their interactions were assessed. Significant findings were then further analyzed in univariate models for the three parameters (change in JTT speed, GPT speed and GPT accuracy). The cumulative learning probability was measured by Kaplan-Meier analysis with Mantel-Cox log rank statistics. Time-related changes in movement time, target error rate, and movement smoothness were assessed by separate repeated measures ANOVAs. Correlations were assessed by Pearson’s r and group differences were compared by Fisher’s z-transformation. Stroke lesion volume and concordance was compared between subgroup of patients by independent t-tests. To take the effect of the BDNF val66met polymorphism on learning into account, total learning per genotype and tDCS condition was compared by four planned independent t-tests (Supplementary Methods). In addition, demographic data (Table 1) were compared between groups by separate one-way ANOVAs. Binary data were compared with Chi-square. Changes in BDI were assessed by repeated measure ANOVA (Table 1, Supplementary Table 1). Anatomical lesion distribution was compared between groups using the VLSM Liebermeister procedure of the MRIcron nonparametrical mapping tool ⁴⁵.

Results

Demographics and psychophysical data

Out of 272 screened patients, 56 patients were enrolled in the study (Fig. 1B). Five patients did not complete the study due to the following reasons: finger infection interfering with task execution (n=1), Botox treatment during the follow-up period altering the motor deficit (n=2), traumatic intracerebral hemorrhage after a fall (n=1), muscle cramping interfering with performing the tests (n=1). One patient (baseline skill >3 SD worse than all others, UEFM 42, severe hand fatigue during training) was excluded. Consequently, data from 50 patients (n= 18 real tDCS, n= 18 sham tDCS, n=14 no-training/no-tDCS) were analyzed. The three groups were comparable in terms of demographical and clinical data (Table 1, Fig. 1C, and Supplementary Table 1). The voxel-wise lesion-symptom mapping analysis did not reveal a group difference in lesion anatomy (p>0.01, FDR corrected). Participants tolerated the tDCS intervention well and were successfully blinded for the type of stimulation (Table 1). They rated the perceived influence of the investigator on their performance similarly neutral to slightly positive in all three groups. Regardless of intervention, patients displayed lower values on the BDI on day 5, suggesting a mood enhancing effect of the study context.

Motor skill learning and generalization

Sessions took place at about 11AM (Table 1). Patients started with comparable SVIPT performance (Table 1; t(34)=0.32, p=0.75, d=0.11). To facilitate visualization of most measures, the learning curve (Fig. 2A) displays change in skill relative to the first block of day 1 (see individual data and non-normalized learning curve in Supplementary Fig. 3).

Overall training effect

Across groups, skill increased significantly over time of training (F(1,34)=167.96, p<0.0001, d=4.45)) and the effect was catalyzed by tDCS, as indicated by a significant group*time interaction (F(1,34)=6.60, p=0.015, d=0.88. There was no effect of group (F(1,34)=0.56, p=0.461, d=0.25).

Total learning (primary outcome)

Total learning was significantly enhanced by tDCS (2.29±0.16 real tDCS versus 1.53±0.25 sham tDCS group; t(34)=−2.57, p=0.02, d=0.86 Fig. 2B, Supplementary Table 2). These values reflect unitless “skill” changes of the log-transformed skill parameter (ln(a)).

Learning stages (online learning, offline learning, retention, rate of forgetting)

Numerical values are presented in Supplementary Table 2. Compared to sham tDCS, patients showed more online learning and less offline learning when stimulated with real tDCS (Fig. 2B) but no statistical differences were found (t(34)=−1.53, p=0.14, d=0.51; t(34)=0.62, p=0.54, d=0.20 respectively). As shown in Figure 2A, most of the skill resulting from training was retained in the follow-ups, regardless of stimulation type. However, there was no specific effect of tDCS on long-term retention assessed at day 113 (Fig. 2B, t(34)=0.58, p=0.56, d=0.16). Accordingly, the time-weighted slope of forgetting over all follow-up time points was not different between groups (t(34)=0.37, p=0.71, d=0.12). The proportion of skill retained on day 113 was similar in both conditions (78± 11% sham vs. 74±10% real tDCS; t(34)=0.29, p=0.77, d=0.10, Fig. 2C). For visual illustration of online and offline learning as well as retention on individual days, see Supplementary Figure 4.

Skill gains remaining at the end of the study

Since real tDCS stimulated patients showed greater total learning, and retention was unaffected, we also found a strong trend for greater remaining skill at the end of the study compared to sham tDCS (t(34)= −1.98, p=0.056, d=0.66).

Generalization

To assess the effects of hand training and tDCS on generalization, we tested patients for improvements in untrained motor tasks (JTT, GPT). This group comparison included the no-training/no-tDCS group. Patients started with similar performances on the two tasks (Table 1; JTT speed: F(2,47)=0.55, p=0.58; GPT speed: F(2,47)=0.21, p=0.82; GPT errors: F(2,47)=0.33, p=0.72). After training, GPT accuracy improved in the two training groups, while being poorer in the no-training/no-tDCS group (Fig. 3A). GPT speed (Fig. 3B) improved in all groups. Training was also associated with greater improvements in JTT speed compared to the no-training/no-tDCS group (Fig 3C). The MANOVA (GPT accuracy, GPT speed, JTT speed) revealed a main effect of time (F(5,30)=6.28, p<0.0001, d=0.81) and training state (F(1,42)=3.203, p=0.033, d=0.58) on generalization, but no interaction (F=0.795, p=0.67, d=0.29). Per univariate analysis, differences related to training state were present in GPT accuracy (F=4.59, p=0.038, d=0.69) and JTT speed (F=4.534, p=0.039, d=0.69), but not GPT speed (F=0.145, p=0.71, d=0.12).

Figure 3. Measures of generalization after SVIPT training and during the follow-up period.

A, Grooved Pegboard test, paretic hand; Improvement measured by absolute change in accuracy relative to day 1; Both training groups show increased accuracy (a reduction in no. of errors is indicated by positive values), on all days, while the no-training/no-tDCS group shows less accuracy (i.e., trading accuracy for higher speed). The MANOVA revealed a significant effect of training compared to no training/no-tDCS, but no additional effect of real tDCS compared to sham tDCS. B, Grooved Pegboard test, paretic hand, Percent improvement in speed (total time to complete the test) relative to day 1. Both training groups and the no-training/no-tDCS group showed improvements in total time. There was no significant difference between trained and untrained patients or between sham and real tDCS stimulated patients. C, Jebsen Taylor test, paretic hand; Percent improvement in speed (total time to complete the full test) relative to day 1. The MANOVA revealed a significant effect of training compared to no- training/no-tDCS, independent of tDCS stimulation type. The no-training/no-tDCS group showed only minor improvements (repetition effects). D, The GPT accuracy change is plotted against the GPT speed change. Data from all time points were used for illustration of the lacking relationship between the two variables in the no-training/no-tDCS group, compared to the strong positive relationship in the two trained groups (indicating improvements in both variables). The real tDCS group shows the greatest improvements in accuracy for a given speed change. The ellipses indicate the 90% confidence interval per group, the lines represents the mean centered linear regression line per group. Significance: *p<0.05. All data except for panel C (single subject data) are shown as group mean ± SEM.

When testing for an additive effect of tDCS, no significant differences between the two tDCS conditions were detected in the MANOVA (group F(1,28)=0.396, p=0.757, d=0.22). As expected, there was a significant effect of time (F(5,42)=6.42, p<0.0001, d=0.87, but no interaction (F=1.562, p=0.192, d=0.43).

Since the separation of GPT performance into two parameters (speed and accuracy) may mask potential changes in the speed-accuracy-tradeoff, we additionally plotted speed and accuracy changes at all time points for visual inspection of their relationship. Fig. 3D suggests a loss of accuracy to achieve higher speeds in the no-training/no-tDCS group, in contrast to the improvement of both parameters in the two trained groups.

Online learning on the first day

For comparison with previous studies commonly using a single session design, we also assessed learning on the first training day, which was significantly greater for real than sham tDCS (t(34)=3.29, p=0.002, d=1.1).

Cumulative learning probability

A greater proportion of patients in the real tDCS group required less training amount to reach a skill gain> 1 compared to sham tDCS (Χ²=8.44; p=0.004, d=1.88, Fig 4A). In other words, proficiency was achieved significantly earlier during training in the real tDCS group. Using the BESD, these data suggest that if treated with real tDCS, the likelihood to reach this particular level of skill increases by 11.1%, corresponding to a number needed to treat of nine patients.

Figure 4. Analysis of cumulative learning probability and correlation of learning subcomponents.

A, Compared to sham tDCS (blue), patients reached a minimum skill gain of 1 unit earlier, when training was combined with real tDCS (red). A skill gain of 1 unit represents the average level of skill reached by the sham tDCS group and was extrapolated from a predictive model. Note that the sham tDCS group is censored, i.e. the Kaplan Meier interval for the time to event (100% of patients reaching the criterion) is not known for this particular group. This happened because few patients did not reach a skill gain of 1 within the 25 training blocks. B, There was a strong negative correlation between online and offline learning during training across groups. Real tDCS (red circles and line) shifted the set point for occurrence of offline skill loss towards higher online learning compared to sham (blue squares and line).

Relation between online and offline learning

Patients gained skill mainly by online learning and subsequently maintained skill between sessions (Fig. 2B). Interestingly, those patients showing the greatest online learning also showed the least offline learning, i.e., there was a strong negative correlation (r=−0.929, p<0.0001, Fig. 4B). Patients receiving real tDCS also showed this negative correlation (r=−0.918, p<0.0001), which was not altered compared to sham (z=−0.205, p=0.419). However, the set point for the occurrence of negative offline learning (forgetting) was shifted towards greater online learning (Fig. 4B).

Changes in movement time, target error rate and movement smoothness

Movement time and target error rate decreased, while movement smoothness increased across training days (F(1,34)=95.433, d=3.35; F(1,34)=43.931, d=2.273; and F(1,34)=28.961, d=1.846 respectively; all p<0.00001, Fig.5). The real tDCS group had significantly faster movement times across sessions compared to sham (group x time interaction: F(1,34)= 8.595, p=0.006, d=1.01), whereas target error rate decreased similarly across sessions with both stimulations (group x time interaction: F(1,34)=0.178, p=0.676, d=0.145). Moreover, real tDCS increased the quality of the movement across sessions, i.e. movement smoothness relative to sham (group × time interaction: F(1,34)=4.167, p=0.049, d=0.7). There were no differences for factor group.

Figure 5. Learning curves for the five training days and the follow-up period for movement time, error rate and movement smoothness.

A, Movement time: both groups showed improvements in movement time; real tDCS significantly shortened movement time across sessions compared to sham tDCS; B, Error rate: both groups reduced the error rate over 5 days; no significant differences were found between real and sham tDCS; C, Movement smoothness: Real tDCS significantly enhanced movement smoothness across sessions compared to sham tDCS. All data are shown as group mean ± SEM.

Correlation of total learning with patient characteristics

Total learning was not influenced by starting performance (R²=0.09; p>0.9), time since stroke (R²=0.21, p>0.9), lesion volume (R²=0.002, p=0.785), lesion location (all t<0.51, p>0.05,) concordance (sham: t(16)= 0.23, p=0.81; real tDCS: t(16)=0.09, p=0.93), or BDNF genotype (sham: (t(16)=0.89, p=0.39, see supplementary results). Across those subjects receiving real tDCS, BDNF met carriers tended to demonstrate less total learning compared to BDNFval/val patients (t(16)=1.93, p=0.072, Supplementary Fig. 5B).

Discussion

In a group of mildly impaired chronic stroke patients, we found that tDCS applied to the ipsilesional M1 during training enhanced learning of the trained skill, predominantly through an effect on the online learning stage. Training was associated with long lasting generalization to untrained upper extremity motor function, with no additional benefit of tDCS.

Stages of visuomotor learning in stroke patients

Chronic stroke patients displayed successful visuomotor learning over multiple days, regardless of stimulation condition. Our findings extend those of single session motor learning studies, which have shown within-session improvements and intact short-term retention of skill ^{11,14,17,46,47}. They also affirm studies finding that patients improve in single performance variables (i.e., movement time) or on speed-accuracy relationships after multi-session motor training^{13,15,16,18,48}. Here, we specifically assessed the stages of motor learning, which had not been previously investigated. We showed that similar to healthy subjects ¹⁰ skill was acquired mainly within session (online) and patients never returned to naïve skill levels between days, indicating motor memory consolidation (i.e. offline maintenance). Interestingly, online and offline skill gains were negatively correlated, i.e. offline skill loss occurred in the patients with greatest online learning. These findings suggest that there may be an intrinsic set-point during the learning process, e.g. by homeostatic cortical network regulation ^49,50, which is difficult to exceed under physiological conditions,. Similar to our findings in healthy subjects¹⁰ and even with less intense training (500 vs. 1000 trials) and a longer follow-up period (4 vs. 3 months), improved skill across patient groups was retained. Collectively, these findings suggest the integrity of the stages of motor learning in chronic stroke patients, despite disordered post-stroke anatomy.

tDCS accelerates and enhances visuomotor learning in chronic stroke

tDCS applied to M1 enhances motor skill learning over multiple days in healthy subjects ^9,10. Similar findings were observed on a first/single training day in elderly subjects ²⁴ and chronic stroke patients ^30,34. Single-session study designs hinder disambiguation between effects of tDCS on performance versus effects on the learning process ^10,51,52. We therefore assessed multi-session online learning, offline learning, and long-term retention. In keeping with our findings in healthy subjects^10,23, total learning was significantly enhanced in real tDCS-stimulated stroke patients compared to sham. A beneficial effect of real tDCS on online learning present on day 1 is consistent with previous studies ^23,30,34, but was not found for the cumulated online learning over 5 days. Since the majority of stroke patients are older adults, it is worth noting that real tDCS-stimulated elderly healthy subjects demonstrate greater day 1 online learning than sham-stimulated subjects, without an effect on offline learning assessed the next day ²⁴. Similarly, in the present study, offline learning over multiple days did not differ between sham and real tDCS patients. These findings are in contrast to our studies in healthy subjects, where real tDCS predominantly enhanced offline learning ^10,23. Hence, there may be a mechanistic difference in how tDCS affects the learning process in the elderly or injured brain, enhancing the acquisition of a skill without augmenting its subsequent consolidation. It is conceivable that healthy subjects saturate online learning through a greater number of training trials or through more proficient learning mechanisms, and real tDCS effects cannot modify this threshold. In the stroke patients, conversely, online learning may remain unsaturated (due to poorer baseline performance, fewer training trials, or less proficient learning mechanisms), resulting in a modifiable process by real tDCS. However, we found an inverse relationship between online and offline learning regardless of stimulation, pointing to a specific set point for the capacity of the motor system to gain skill. Real tDCS slightly altered this relationship compared to sham, such that for given online gains, there was marginally less offline loss. It is thus tempting to speculate that offline learning processes may be secondarily influenced as a consequence of a specific effect of tDCS on online learning, e.g. through metaplastic effects⁴⁹. As in healthy subjects¹⁰, retention of acquired skill was not affected by real tDCS. It is thus important to note that any long-lasting benefit of real tDCS (i.e. skill remaining at day 113) is derived from the primary effect in the training period. While this is the only study on long-term skill retention extending several months, similar findings are visible in previous studies that suggest a lack of effect of tDCS on short-term retention ^30–32,34.

Training, but not tDCS, improves generalization of visuomotor learning

To be considered a valuable interventional tool in neurorehabilitation, tDCS combined with training should ideally improve motor function beyond its augmentation of trained skills. Here, we observed generalization after SVIPT training, evidenced by enhanced accuracy on the GPT and faster execution speed in the JTT. Such improvements were not found in the non-trained, non-stimulated patients. This third group enabled us to disambiguate learning effects resulting from repetition of the probe tasks from generalization of effects from training. We found no additional benefit of tDCS on generalization compared to sham. A recent meta-analysis revealed an overall beneficial effect of tDCS in combination with various motor training/therapy concepts in subacute and chronic stroke patients ⁵³. This effect was evident as improved performance on either clinical scales (e.g. UEFM, NIHSS) or motor-specific task improvements (9-hole peg test, visuomotor tracking). Allman and colleagues reported improved clinical scores resulting from real tDCS combined with bimanual motor training, evident for at least 3 months after training ²⁵. One group has also found enhanced generalization on a pegboard task one week after a single session of bihemispheric tDCS and unimanual visuomotor training ³⁰ (but findings were not replicable in a similar study design ³²). It is difficult to interpret such improvements as motor generalization, since these studies lacked non-trained and non-stimulated control groups. Our study extends these findings, showing robustness of training-related generalization for four months after the end of training, but no additive effect of tDCS.

tDCS-improved skills result from a speed-based shift in the speed-accuracy tradeoff

The majority of previous tDCS studies in stroke patients did not specifically address the speed-accuracy tradeoff ^25,34,46,48. In the SVIPT, healthy subjects and chronic stroke patients (performing with the elbow flexors) typically improve skill through a reduction in movement time, while also improving or maintaining a constant error rate^18,54,55. Here, our sham tDCS patients improved skill on the SVIPT through reductions in both movement time and error rate. Relative to sham, the real tDCS group further reduced movement times, without differences in error rate. Hence, tDCS led to a predominantly speed-based (leftward) shift in the speed-accuracy-tradeoff, extending on a similar finding following a single training session in another visuomotor task³⁰. Since movement smoothness significantly increased with real tDCS, it is conceivable that better movement control contributed to the ability to perform the task at higher speeds without sacrificing accuracy. We consider this further evidence that tDCS does not solely affect a single component of movement (e.g. speed) but rather changes the relation between different movement components. Depending on the task, this may be reflected differently in the speed-accuracy parameter space. Similar to the SVIPT, training-related improvements in the GPT were due to greater accuracy changes at similar to slightly faster movement times compared to the no training/no tDCS group. However, tDCS did not affect this speed-accuracy-tradeoff, providing further evidence for the restriction of the tDCS effect to the trained skill.

Limitations and clinical implications

We found evidence for a task-specific training-enhancing effect of tDCS with the anode applied over the ipsilesional primary motor cortex in chronic stroke patients. Because this study was planned and powered to address cumulative effects of tDCS on total learning as well as online/offline effects across 5 days of training, we are unable to assess stimulation effects on individual training days. Given the ostensible lack of focality of tDCS ⁵⁶, we are also unable to infer which stimulated cortical areas may contribute most to learning. Since our study required that patients have sufficient hand function to execute the SVIPT, it is unclear how our results would translate to more severely affected patients. We observed long-lasting generalization to untrained upper extremity function due to training, but no additional benefit provided by tDCS. While this limitation may relate to our choice of well-recovered patients and our focus on skilled tasks, future studies could include a more impaired patient cohort and more multifaceted generalization measures to comprehensively gauge stimulation effects in the clinical context. Moreover, it is important to investigate these effects in the acute stage of stroke, when tDCS-augmented activity-dependent plasticity could positively interact with injury-induced plasticity.