Uncovering the Role of the Human Hippocampus in Procedural Motor Learning: Insights from Implicit Sensorimotor Adaptation

Abstract

Recent evidence suggests that the human hippocampus, traditionally associated with declarative memory, plays a role in motor sequence learning (MSL). However, the classic MSL paradigm depends initially on declarative learning. Thus, it is critical to discern whether the participation of the hippocampus relates to its canonical role or to processing a general aspect of learning that transcends the declarative/non-declarative distinction. To address this issue, here we turned to visuomotor adaptation (VMA)—a type of motor learning involving skill recalibration—which unlike MSL can be easily manipulated to eliminate the explicit component. We examined the broader involvement of the hippocampus in procedural motor learning by using diffusion MRI in a sample of males and females to assess structural plasticity associated with memory consolidation in VMA and an implicit-only version (IVMA). We found that both VMA and IVMA engaged the left posterior hippocampus in a learning-specific manner. Remarkably, while VMA induced only transient hippocampal alterations, IVMA elicited structural changes that persisted overnight, underscoring the reliance on implicit learning for enduring neuroplasticity. As expected, training on both tasks impacted the microstructure of the cerebellum and the motor and posterior parietal cortex. Notably, the temporal dynamics of changes in these regions paralleled those of the left hippocampus, suggesting that motor and limbic regions operate together as part of the same network. Collectively, our findings support an active role of the hippocampus in implicit motor learning and argue for a unified function in memory encoding regardless of the declarative or non-declarative nature of the task.

Significance Statement

Brenda Milner's work on Patient H.M. introduced the concept of specialized memory systems in the brain but also inevitably initiated a dichotomy in cognitive neuroscience, with episodic learning viewed as hippocampus-dependent and procedural learning as hippocampus-independent. This distinction fragmented the field of cognitive neuroscience, with studies on declarative and procedural memory progressing along parallel paths. Here, we show that a purely implicit motor task induces learning-specific structural plasticity in the human hippocampus. Notably, the temporal dynamics of these hippocampal changes were mirrored by key motor regions involved in acquisition, pointing to a close interaction between memory systems. Our findings provide evidence supporting the involvement of the human hippocampus in procedural motor memory and argue for common mechanisms supporting memory formation.

Introduction

Brenda Milner's seminal work on patient H.M. revealed that the surgical bilateral resection of the hippocampus markedly impaired the encoding of declarative memory (Scoville and Milner, 1957). Nevertheless, H.M. maintained the capacity to learn non-declarative motor tasks (Milner, 2005). Milner introduced the concept of specialized memory systems in the brain, which initiated a dichotomy in cognitive neuroscience, with episodic learning viewed as hippocampus-dependent and procedural learning as hippocampus-independent.

Recent neuropsychological and neuroimaging studies have challenged this traditional view, prompting a reassessment of the hippocampus's role in procedural memory (Albouy et al., 2013; Deleglise et al., 2023). Using the motor sequence learning (MSL) paradigm involving executing a five-item finger sequence with the non-dominant hand (Karni et al., 1995), it has been shown that hippocampal dysfunction in amnesic patients impairs both the rate of learning and memory consolidation (Döhring et al., 2017; Schapiro et al., 2019). Similar outcomes have been observed in epileptic patients with hippocampal damage (Long et al., 2018). Thus, although hippocampal dysfunction may not hinder motor learning to the same extent as episodic learning, it has a measurable impact on the efficacy of memory encoding/consolidation.

In line with the neuropsychological evidence, we have demonstrated that MSL improvements in performance observed during the rest periods interleaved with practice (micro-offline gains) are linked with an increase in hippocampal activity (Jacobacci et al., 2020a). This finding is reminiscent of neural replay, long associated with memory reactivation in rodents (Foster, 2017; Buch et al., 2021). Functional changes were followed by rapid (∼30 min) alterations in hippocampal microstructure evidenced by a reduction in mean diffusivity (MD), a metric derived from the diffusion tensor imaging (DTI; Basser et al., 1994). Decreases in MD during learning are associated with synaptogenesis and astroglial expansion, two well established markers of gray-matter structural plasticity linked to LTP-like processes (Holtmaat and Svoboda, 2009; Sagi et al., 2012). Thus, we have proposed that the same neuronal ensembles that reactivate during the quiet rest periods of MSL may undergo structural plasticity, a mechanism in line with the modern definition of an engram (Josselyn and Tonegawa, 2020). Recent neurophysiological work showing evidence of neural replay in the human hippocampus associated with micro-offline gains (Chen et al., 2024; Sjøgård et al., 2024) offers preliminary support for this hypothesis.

Collectively, these studies suggest that beyond its role in declarative learning, the human hippocampus also supports procedural motor learning. However, the classic MSL paradigm depends initially on declarative learning as participants must memorize the sequence. Thus, the participation of the hippocampus may be linked to its canonical role in processing the declarative component of the task. While MSL cannot be modified to eliminate explicit learning without substantially altering the paradigm (Krakauer et al., 2019), visuomotor adaptation (VMA)—a form of skill recalibration in response to visual perturbations—can be easily adapted to isolate implicit learning (Taylor et al., 2014; Morehead et al., 2017; Kim et al., 2018). This makes it an ideal experimental paradigm for studying the broader involvement of the hippocampus in procedural memory.

In this study, we investigated whether the human hippocampus is involved in implicit motor learning by using MD to track structural plasticity induced by learning on the VMA paradigm. Building on our previous longitudinal design, we measured MD in the classic VMA paradigm and its implicit-only version (IVMA) in the short (30 min) and long (24 h) timescales. We predicted that if the hippocampus is indeed required for procedural motor learning, both VMA and IVMA would induce structural changes in this brain region. Additionally, based on the competition hypothesis, which suggests that implicit and explicit learning compete for error during VMA (Albert et al., 2022), we hypothesized that changes in hippocampal microstructure induced by IVMA would be more persistent than those induced by VMA (Sagi et al., 2012). Finally, we explored the interaction between the hippocampus and the motor system in supporting motor learning.

Materials and Methods

Participants

A total of 40 participants aged between 18 and 34 years old (mean ± SD = 24.8 ± 3.9 years old; 26 females) were enrolled in the study. All participants were healthy volunteers with no self-reported history of psychiatric, neurological, or cognitive disorders, nor any history of sleep disturbances. Participants were instructed to abstain from alcohol on the day before and during the experiment, as well as to maintain their regular sleep habits. All subjects were right-handed as assessed by the Edinburgh handedness inventory (Oldfield, 1971). Written consent was obtained from all participants, and they were remunerated for their time and inconvenience. The experimental procedure was approved by the local Ethics Committee (University of Buenos Aires) and was conducted following the Declaration of Helsinki.

Experimental paradigms

VMA

The classic VMA paradigm used here has been described in previous studies (Villalta et al., 2015; Lerner et al., 2020; Solano et al., 2022) and is summarized in this section. This center-out task requires participants to hit one of eight visual targets displayed concentrically around the start position and equidistantly on a computer screen using a joystick controlled with the thumb and index finger of their right-dominant hand (Fig. 1A). In this paradigm, the vision of the hand is occluded. On each trial, subjects executed a shooting movement to one of the eight targets presented individually on the screen in a pseudorandomized order. One cycle consisted of eight trials directed to each target; 11 cycles composed a block. Continuous visual feedback on the hand's position was provided through the cursor from the onset of each trial until the movement was completed. To avoid online corrections that would lead to submovements, the joystick's gain was set to 1.4 so that a displacement of 1 cm of the tip of the joystick moved the cursor on the screen by 1.4 cm. According to previous pilot data from our lab, this gain yields straight paths with little or no online corrections (Villalta et al., 2015).

Figure 1.

Experimental paradigms and experimental design. A, Experimental paradigms. Subjects trained on one of the two VMA paradigms, consisting of making center-out movements to one of eight visual targets using a joystick while the vision of their hand was occluded. The inset illustrates the visual display of the computer screen across different stages of learning: baseline, early, and late adaptation to an imposed visual perturbation. In the classic paradigm known as VMA, the cursor represents the direction of the hand movement. Subjects first performed a set of null trials for familiarization with no perturbation (baseline). Then, an abrupt visual perturbation in the form of an optical rotation (40° clockwise) was applied to the cursor representing the hand (early adaptation). With time, subjects learned to compensate for the imposed visual rotation (late adaptation) as indicated by the pointing angle (α; i.e., the angle delimited by the straight line connecting the start point with the target position and the movement direction of the hand). In the IVMA task, after the baseline period, the cursor's trajectory was clamped to 40° clockwise relative to the target's location throughout the experiment (early adaptation). In this manipulation the cursor did not match the movement direction of the hand. Decoupling the cursor from the hand's movement direction avoids improvements in performance based on explicit strategy, forcing subjects to rely on the implicit system. Despite not seeing their trajectories, subjects learn to compensate for the optical rotation (late adaptation). B, Experimental design. Two groups of subjects participated in this study. One group (n = 18) trained on the VMA task, while the other group (n = 22) trained on the IVMA paradigm. Diffusion MRI (dMRI)—and T1w images—were acquired before (baseline), 30 min, and 24 h after learning to track the dynamics of changes in microstructure induced by VMA and IVMA in the hippocampus and key motor regions. Overnight memory retention (test) was assessed using two EC trials. The test session took place after the MRI acquisition obtained 24 h post-training.

The VMA paradigm involved three different types of trials. During null trials, the cursor movement directly mapped the joystick's movement in native coordinates. During perturbed trials, a CW 40° optical rotation was imposed on the cursor, deviating its trajectory. During error-clamp (EC) trials, the cursor trajectory was artificially manipulated to simulate “straight” paths to the target, resembling that of correct trials, by projecting the actual cursor's movement onto a straight line with additional variability (0 ± 10°, mean ± standard deviation). EC trials allowed estimating memory retention by eliminating learning from error (Criscimagna-Hemminger and Shadmehr, 2008).

Implicit VMA

Two main learning processes contribute to VMA, a fast explicit learning process that involves compensating the optical rotation using a conscious aiming strategy and a slow implicit learning process involving automatic recalibration (Taylor et al., 2014; Albert et al., 2022; Tsay et al., 2022). The former is thought to be driven by errors in achieving the movement goal (task error), whereas the latter is thought to be driven by errors in predicting the sensory outcome of a movement (sensory prediction error). To study implicit adaptation in isolation of the explicit component, we fixed task error by clamping the cursor direction to 40° CW relative to the target's position throughout training (Fig. 1A, bottom). In this manipulation, the cursor matches the speed of the hand while maintaining no contingency with the hand angle, ensuring that learning is driven by the implicit component (sensory prediction error; Morehead et al., 2017). Despite their awareness of the manipulation, participants adapt to the perturbation implicitly (Kim et al., 2018). We refer to this task as IVMA.

As in the VMA paradigm, each cycle consisted of eight trials, and each block consisted of 11 cycles. The IVMA paradigm involved three different types of trials. During null trials, the cursor movement directly mapped the joystick's movement. During clamped-to-40° trials, a fixed CW 40° visual rotation was imposed on the cursor relative to the visual target's position. Finally, during clamped-to-0° trials, the cursor trajectory was fixed to the visual target position. The latter was used to quantify memory retention in this paradigm because it avoids further learning while providing no veridical feedback on hand movement.

Both VMA and IVMA tasks were implemented in MATLAB (The MathWorks) using the Psychophysics Toolbox v3 (Brainard, 1997).

Experimental design and procedure

This study was designed to examine the involvement of the hippocampus in implicit motor learning. To this aim, we used MD, an indirect marker of structural plasticity, to quantify changes in microstructure induced by the classic VMA paradigm—relying both on explicit and implicit learning—and by IVMA, in the short and the long term. To be able to contrast the VMA tasks with the MSL task used in our previous study (Jacobacci et al., 2020a), T1w (T1-weighted) and dMRI images were acquired following the same longitudinal design: 30 min before and 24 h after learning (Fig. 1B). Overnight memory retention was assessed 24 h after training.

Two groups of subjects participated in this study. One group (n = 18; mean ± SD = 25.3 ± 4.2 years old; 12 females) trained on the VMA task, while the other group (n = 22; mean ± SD = 24.7 ± 3.8 years old; 14 females) trained on the IVMA paradigm. Both groups were instructed to perform a shooting movement to one of the eight targets with the cursor as soon as it appeared on the screen. Before practice, subjects became familiar with the setup by performing a few center-out movements with veridical feedback during null trials. Additionally, participants in the IVMA group were told of the clamped-cursor manipulation and were instructed to ignore the cursor direction and aim directly at the target throughout the experiment. Participants in the VMA group performed one block of null trials followed by five blocks of perturbed trials, while participants in the IVMA group performed one block of clamped-to-0° trials followed by five blocks of clamped-to-40° trials. Training for each task lasted ∼25 min.

Data acquisition and preprocessing

Magnetic resonance (MR) images

MR images were acquired using a 3 T Siemens Tim TRIO MRI scanner located at the Instituto Angel Roffo, University of Buenos Aires, equipped with a 12-channel head coil. MRI images were acquired using a longitudinal design before (baseline), 30 min, and 24 h post-training on VMA or IVMA, as indicated in Figure 1B. On each session, an anatomical T1w MPRAGE image was collected with the following parameters: repetition time (TR) = 2,530 ms; echo time (TE) = 2.17 ms; inversion time (TI) = 1,100 ms; flip angle (FA) = 7°; bandwidth (BW) = 199 Hz/Px; field of view (FOV) = 256 × 256 mm²; acquisition matrix = 256 × 256; 1 mm isotropic voxels; slices = 176; parallel acquisition, GRAPPA mode; and acceleration factor = 2. The acquisition was performed in the sagittal plane.

Three dMRI acquisitions were collected on each session using the SE-EPI sequence with the following parameters: two dMRIs were acquired with an anterior–posterior (A–P) phase encoding direction and one with the opposite direction (P–A); multiband acceleration factor = 2 (Uğurbil et al., 2013; Xu et al., 2013); TR = 5,660 ms; TE = 89 ms; FA = 90°; BW = 1,488 Hz/Px; slices = 76; 2 mm isotropic voxels; FOV = 240 × 240 mm²; EPI factor = 120; acquisition, interleaved; and 30 monopolar gradient directions with b value = 1,000 s/mm². Furthermore, seven b0s were acquired using the same phase encoding direction: two at the beginning of the acquisition, one at the end, and the rest interleaved every five b − 1,000 volumes.

Preprocessing and normalization of the dMRIs were performed using the FMRIB Software Library (FSL; University of Oxford; version 6.0.3) and the Advanced Normalization Tools (ANTs; Wellcome Department, UCL; version 2.4.2). The preprocessing steps were conducted separately for each scanning session and included the correction of susceptibility-induced distortions using FSL's Topup (Andersson et al., 2003) and correction of eddy current-induced distortions, head motion correction, and b-vector rotation using FSL's eddy tool (Andersson and Sotiropoulos, 2016). Subsequently, FSL's DTIfit was used to fit a diffusion tensor model and produce the scalar maps for FA and MD metrics. Following preprocessing, DTI scalar maps were normalized to MNI152 stereotaxic space using FA as an intermediate template. We used ANTs for normalization in a pipeline created by our group to minimize across-session test–retest reproducibility error (Jacobacci et al., 2020b). Finally, normalized DTI measures were smoothed with FSL's smoothing tool with a 4 mm full-width at half-maximum Gaussian kernel.

Hippocampal subfields

Human hippocampal subfields were obtained using FreeSurfer (version 7.0). The FSL's MNI T1 brain image of 1 mm resolution was processed through FreeSurfer's recon-all function. Subsequently, we used the segmentHA_T1 optimized pipeline to extract the hippocampal subfields. This specific pipeline allowed segmenting the hippocampus along its longitudinal axis into the head, body, and tail (Iglesias et al., 2015).

Data analysis and statistics

Behavior

Motor performance for both VMA and IVMA tasks was quantified based on the pointing angle, that is, the angle determined by the movement direction of the joystick and the line segment connecting the start point with the target position (Lerner et al., 2020; Solano et al., 2022). Trials in which the pointing angle exceeded 120° were excluded from subsequent analysis. Given that trials were organized into cycles, trial-by-trial data were transformed into cycle-by-cycle time series by computing the median pointing angle for each cycle and subject. The asymptotic performance for each participant was determined by calculating the mean of the last five cycles during the final adaptation block.

To assess memory retention for each task and each subject, we used two EC cycles for VMA and two clamp-to-0° cycles for IVMA, relative to each subject's asymptotic performance according to this equation: [PA_A − abs(PA_A − PA_{EC/clamp-to-0})] ∗ 100/PA_A, where PA_A is the pointing angle of the asymptote and PA_{EC/clamp-to-0} is the pointing angle of the two retention cycles. Finally, this percentage measure was averaged across the two cycles. The rate of learning was assessed by calculating the median of the pointing angle across the first four cycles—excluding the first one—during the adaptation phase (Huberdeau et al., 2015; Kim et al., 2018) and expressed as a percentage of the pointing angle asymptote. Three subjects from the IVMA group were excluded from the behavioral analysis because their pointing angle of the asymptote was <0°.

dMRIs

Longitudinal changes in MD induced by VMA and IVMA across sessions were computed using a nonparametric sandwich estimator (SwE; Guillaume et al., 2014) model with a permutation-based threshold–free cluster enhancement (TFCE) approach (Smith and Nichols, 2009). SwE has been specifically designed for accurate modeling of longitudinal and repeated-measure neuroimaging data (version 2.0.0). To detect hippocampal changes in microstructure across sessions (baseline, 30 min, 24 h), statistical analyses were conducted separately for each task within a bilateral hippocampal mask derived from the Harvard-Oxford subcortical structural probability atlas (>30% probability, FSL). To compare changes induced by VMA versus IVMA, we conducted a task (VMA, IVMA) by session (baseline, 30 min, 24 h) interaction analysis within the same hippocampal mask. Furthermore, to identify changes in microstructure at the level of the motor system, we conducted a separate whole-brain SwE analysis across sessions as implemented in our previous study (Jacobacci et al., 2020a) for VMA and IVMA and their interaction. All nonparametric SwE analyses were conducted based on 5,000 permutations, and significant clusters were identified using a family-wise error (FWE)-corrected p value <0.05.

Following cluster identification, MD values were extracted from all voxels within each cluster for each subject and session, and the median MD was computed. MD changes for each cluster at 30 min and 24 h were expressed relative to the baseline [e.g., (30 min MD − baseline MD) / baseline MD]. For each cluster and each session, we computed the mean percentage change in MD across subjects and determined the 95% confidence intervals (CI). This was achieved using the summarySEwithin function from the Rmisc package in R (R Core Team, 2020). This function is designed to normalize within-subject data effectively by removing between-subject variability using the method proposed by Morey et al. (2008).

To examine whether motor regions identified in the whole-brain analysis displayed temporal dynamics similar to those of the left hippocampus—consistent with the existence of a shared network spanning limbic and motor regions—we conducted separate linear mixed models (LMMs) for VMA and IVMA. These models included random intercepts for each subject to account for repeated measures. The primary variable of interest was MD change, with fixed factors being the regions of interest (ROIs) drawn from significant clusters of the whole-brain analysis (hippocampal and motor regions) and session (30 min and 24 h). A session-by-region interaction was tested to assess whether MD changes followed a parallel trajectory across regions or displayed distinct temporal dynamics. Additionally, to assess whether MD changes across these brain regions were not independent but likely influenced by a shared underlying factor, we conducted a complementary principal component analysis (PCA) for VMA and for IVMA groups.

To anatomically compare structural changes induced by different motor learning tasks, we conducted the interaction between the MSL and the active control condition (CTL task) from our previous study (Jacobacci et al., 2020a), within the bilateral hippocampal mask using the same SwE approach, and overlaid these results from the ones obtained in the current study.

In our prior MSL study, we found that rapid changes in MD related to gains in performance during the early stages of learning. To evaluate if changes in MD induced by VMA and IVMA were also specific to memory encoding, we computed a Pearson's correlation (R Core Team, 2013) between the percentage change in MD observed 30 min post-learning and the rate of learning.

Of note, TFCE parameters for all SwE analyses within the hippocampal mask were set to H = 4.5 and E = 0.5, to conform to a more conservative approach, whereas those for the whole-brain analysis were set to relatively less conservative parameters to ensure good sensitivity (H = 4 and E=0.5). Note, however, that the chosen TFCE parameters were always more conservative than the parameter settings recommended by Smith and Nichols (H = 2 and E = 0.5).

Results

VMA and IVMA induce structural changes in the human hippocampus consistent with neuroplasticity

Traditionally, the human hippocampus has been associated primarily with declarative memory. However, recent studies from our group and others have reported its involvement in motor learning, suggesting an expanded role beyond its canonical function to non-declarative, procedural memory (Schendan et al., 2003; Gheysen et al., 2010; Albouy et al., 2013, 2015; Döhring et al., 2017; Long et al., 2018; Schapiro et al., 2019; Jacobacci et al., 2020a; Deleglise et al., 2023). Notably, the bulk of the evidence linking the hippocampus to motor learning, including our own (Jacobacci et al., 2020a), originates from studies using the MSL task, which depends initially on declarative learning as participants must first memorize the sequence (Robertson et al., 2004; Krakauer et al., 2019). Thus, a fundamental question in neuroscience is to discern whether the participation of the hippocampus in motor learning relates to its canonical declarative role or to processing a general aspect of learning that transcends the declarative/non-declarative distinction.

To address this question, we turned to the classic VMA paradigm, which, unlike MSL, can be easily modified to eliminate the explicit component. Previous studies have demonstrated that explicit learning can be eliminated by clamping the cursor to a fixed movement direction, thereby making task errors constant (Taylor et al., 2014; Morehead et al., 2017). This decoupling forces learning to rely solely on sensory prediction errors, resulting in a fully implicit version of the paradigm (IVMA; Fig. 1A, bottom).

Two groups of participants trained on either VMA (n = 18) or IVMA (n = 22), and we examined the impact of motor learning on the temporal dynamics of MD in both the short and long term using DTI (Fig. 1B). Diffusion MRI images were acquired at before learning (baseline), 30 min, and 24 h post-training. Evidence from our group and others suggests that sensorimotor memories consolidate within a 4 to 6 h window (Walker et al., 2003; Lerner et al., 2020; Solano et al., 2024). Thus, our longitudinal design allowed us to sample brain activity early during consolidation and after memory stabilization (Lerner et al., 2020; Solano et al., 2024).

Both groups achieved asymptotic performance after practice on the VMA and IVMA tasks, as reflected by the pointing angle (Fig. 2A,B, top), and showed similar levels of overnight memory retention assessed using EC trials [VMA, mean ± standard error (SE), 52.18 ± 4.33%; IVMA, 49.43 ± 4.79% relative to the asymptote].

Figure 2.

VMA and IVMA induce structural changes in the human hippocampus consistent with neuroplasticity. A, The VMA learning curve (leftmost plot), depicting the pointing angle (mean ± SE) as a function of movement cycles (1 cycle = 8 trials), and memory retention (middle plot) assessed 24 h post-training using EC trials (barplots, mean ± SE relative to the asymptote) are shown. The bottom figure depicts the results of applying the SwE statistical model on MD within a bilateral hippocampal mask (depicted in pink in the brain sections) across the three dMRI sessions; the corresponding barplots quantify the mean percentage change of MD relative to the baseline ± 95% CI for the identified cluster at 30 min and 24 h post-training. VMA induced a significant decrease in MD over the LH 30 min postlearning that reverted to baseline levels by 24 h (p < 0.05 FWE-corrected). The rightmost plot illustrates the Pearson's correlation between the rate of learning and the percentage decrease in MD observed 30 min post-training (R = −0.55; p = 0.018). B, The learning curve for IVMA (leftmost plot) and memory retention (middle plot) assessed 24 h post-training using clamp-to-0° trials are shown. IVMA induced a significant decrease in MD (bottom figure) over the left (bottom inset) and right (top inset) hippocampi 30 min post-learning; however, only structural changes in the LH persisted 24 h post-training (SwE, p < 0.05 FWE-corrected). The rightmost plot shows the Pearson's correlation between the rate of learning and the percentage decrease in MD observed over the LH 30 min post-training (R = −0.51; p = 0.025). C, A task-by-session interaction confirmed the different temporal dynamics of MD for VMA and IVMA (SwE, p < 0.05 FWE-corrected; LH, bottom inset; RH, top inset). Transient changes in microstructure are illustrated by light gray barplots, while those that persisted overnight are shown in dark gray. D, Hippocampal clusters identified for VMA and IVMA are rendered in a diagram of a bilateral 3D hippocampal mask, segmented along its longitudinal axis into the head, body, and tail (Iglesias et al., 2015). Structural changes induced by VMA are illustrated in yellow, while those induced by IVMA are in pink. The results from our previous study showing structural changes induced by MSL are overlaid in blue. RH: right hippocampus; LH: left hippocampus; SwE: sandwich estimator.

Building on our MSL findings (Jacobacci et al., 2020a), we initially conducted a voxelwise ROI analysis of MD within a bilateral hippocampal mask to identify clusters modulated by motor learning. In line with our previous study, we found that VMA induced a reduction in MD over the left hippocampus 30 min post-learning, which reverted to baseline levels by 24 h (SwE, F_(2,17) = 15.84; p = 0.001; Fig. 2A, bottom). In contrast, training on IVMA resulted in a reduction in MD over the same area that persisted overnight, i.e., 24 h post-learning (SwE, F_(2,21) = 20.61; p = 0.001; Fig. 2B, bottom). Notably, IVMA also induced a transient but more subtle reduction in MD over the right hippocampus, which reverted to the baseline by 24 h (SwE, F_(2,21) = 12.29; p = 0.017; Fig. 2B, bottom). The differing temporal dynamics of MD observed for VMA and IVMA are confirmed by the group analysis results shown in Figure 2C (task-by-session interaction, SwE, left hippocampus, F_(1,34) = 7.74; p = 0.001; right hippocampus, F_(1,34) = 8.91; p = 0.001).

Interestingly, the magnitude of the reduction in MD observed 30 min post-training over the left hippocampus correlated with the rate of adaptation in both groups (VMA, Pearson's correlation, R = −0.55; p = 0.018; Fig. 2A; IVMA, Pearson's correlation, R = −0.51; p = 0.025; Fig. 2B), indicating that changes in hippocampal microstructure were likely related to learning (although it is worth noting that the relationship did not reach significance for VMA in the Spearman correlation analysis, suggesting a weaker association compared with IVMA; R = −0.3; p = 0.28 for VMA; R = −0.5; p = 0.03 for IVMA).

Finally, to examine the topography of structural changes induced by motor learning, we overlaid the results obtained from training on VMA, IVMA, and MSL (from our previous study) on the bilateral hippocampal mask segmented into the head, body, and tail (Iglesias et al., 2015). As depicted in Figure 2D, we found that the three motor learning paradigms induced consistent changes over the posterior portion of the left hippocampus, with a clear anatomical overlap over the hippocampal body. Notably, structural changes induced by IVMA were more extensive, spanning both the body and tail of the left and right hippocampi. Interestingly, neither of the motor tasks recruited the anterior subfield, suggesting a consistent involvement of the posterior hippocampus in motor learning.

Collectively, our results provide evidence supporting the role of the hippocampus in implicit motor learning. The overlapping spatial topography of MD changes observed across motor paradigms points to an involvement of the left posterior hippocampus in sensorimotor adaptation. Contrary to the prediction based on the hippocampus's canonical role in declarative learning, the suppression of explicit strategies in IVMA resulted in enduring structural changes. This finding is consistent with the biological implications of a computational model proposing that implicit and explicit learning compete for the available error (Albert et al., 2022), potentially impacting on the neural substrates of memory consolidation.

During adaptation motor and limbic regions operate together as part of the same network

Abundant experimental evidence underscores the involvement of the cerebellum, the posterior parietal cortex (PPC), and the primary motor cortex as key motor regions involved in VMA. Specifically, it has been shown that the cerebellum and PPC are recruited during early and late phases of acquisition, respectively (Della-Maggiore et al., 2004; Tseng et al., 2007; Criscimagna-Hemminger et al., 2010; Galea et al., 2011; Izawa et al., 2012), whereas the primary motor cortex has been linked to memory consolidation and storage (Richardson et al., 2006; Hadipour-Niktarash et al., 2007; Galea et al., 2011; Landi et al., 2011; Orban de Xivry et al., 2011; Villalta et al., 2015).

Thus, we conducted whole-brain analyses to investigate whether VMA and IVMA induces structural changes consistent with neuroplasticity in motor regions and whether the dynamics of these changes align with those observed in the hippocampus. We found that VMA induced a transient (Fig. 3A) decrease in MD 30 min post-learning over the left lateral cerebellum (Lobule VI and Crus I; Grodd et al., 2001; Donchin et al., 2012) and the left PPC, both of which followed the temporal dynamics of MD in the left hippocampus (SwE, p < 0.05 FWE-corrected). IVMA also reduced MD in the same network (Fig. 3B), but these changes persisted overnight, mirroring the pattern observed in the left hippocampus. In addition, IVMA also induced persistent changes in the contralateral primary motor cortex (Boling et al., 1999) and the right lateral cerebellum, as well as a transient decrease in MD over the right hippocampus (SwE, p < 0.05 FWE-corrected). Structural differences across tasks were confirmed by a task-by-session interaction (SwE, p < 0.05 FWE-corrected).

Figure 3.

During adaptation motor and limbic regions operate together as part of the same network. A, A whole-brain SwE analysis conducted across sessions identified structural changes induced by VMA in the left lateral cerebellum, left PPC, and the same region of the posterior hippocampus detected within the hippocampal mask. In all regions, VMA induced a decrease in MD 30 min postlearning that reverted to baseline values by 24 h. B, A whole-brain SwE analysis identified structural changes induced by IVMA over the same brain regions and the primary motor cortex, but in this case, the reduction in MD persisted overnight. IVMA also induced transient structural changes over the right hippocampus. Barplots show the mean percentage change ±95% CI for the time course of MD in these clusters. Transient changes in microstructure are illustrated in light gray, while those that persisted overnight are depicted in dark gray.

To examine further whether limbic and motor regions function as components of a common network, we conducted two complementary analyses. Specifically, to assess whether the hippocampus and key motor regions exhibit a similar temporal dynamic of MD, we performed a separate LMM for VMA and IVMA conditions. Each model incorporated MD changes from both sessions (30 min and 24 h) and all ROI identified in our whole-brain analyses (Fig. 3A,B) as fixed effects, with subjects modeled as random effects to account for repeated measures. For VMA, the analysis revealed a significant intercept (t₍₁₈₎ = −3.21; p = 0.002), demonstrating an overall MD reduction across all examined brain regions (left hippocampus, cerebellum, and PPC). Consistent with the transient MD decrease depicted in Figure 3A, we found a significant main effect of session (F_(1,85) = 46.46; p < 0.001). Crucially, the session-by-region interaction was nonsignificant (F_(2,85) = 1.14; p = 0.33), indicating that the left hippocampus and key motor regions exhibited comparable temporal patterns of MD changes. Similarly, for IVMA, the analysis yielded a significant intercept (t₍₂₂₎ = −2.99; p = 0.003), indicating a global MD decrease. Importantly, we found no significant session-by-region interaction (F_(4,189) = 0.30; p = 0.88), demonstrating that MD changes followed similar temporal dynamics across the hippocampus, cerebellum, motor cortex, and PPC. This synchronized microstructural plasticity across diverse regions provides compelling evidence for their operation as components of an integrated network during IVMA memory consolidation. Additionally, across all regions, we observed a significant main effect of session (F_(1,189) = 9.29; p = 0.003), indicating that despite their persistence, these coordinated network changes continued to evolve during the 24 h consolidation period.

As a second approach to assess whether MD changes across these regions are not independent but are likely influenced by a shared underlying factor, we conducted a separate PCA for VMA and IVMA groups. In the VMA group, the first principal component (PC1, weights for VMA, left cerebellum, 0.6; left hippocampus, 0.56; left PPC, 0.57) explained 69% of the variance (PC2 = 18%; PC3 = 13%), whereas in the IVMA group, PC1 (weights for IVMA, left hippocampus, 0.64; left parietal, 0.62; left cerebellum, 0.46) accounted for 57% of the variance (PC2 = 27%; PC3 = 16%). The pattern of weights identified in PC1 across groups suggests that MD varies coordinately in these regions in both groups, consistent with a distributed network.

Collectively, our findings suggest that during adaptation, motor and limbic regions function together as part of the same neural network. The distinct temporal dynamics observed across tasks open the possibility that, depending on the reliance on implicit learning, this network may experience either short- or long-lasting microstructural changes. These changes could reflect different biological processes involved in neuroplasticity, ranging from rapid homeostatic adjustments during the initial stages of synaptic potentiation to slower structural remodeling in late stages.

Finally, it is important to note that the whole-brain analysis reproduced the anatomical localization and temporal dynamics of hippocampal changes identified with the bilateral mask, further strengthening our findings.

Discussion

Traditionally, the human hippocampus has been primarily associated with the encoding of declarative memory. In this study, we present compelling evidence from an implicit motor learning paradigm suggesting an expanded role to procedural motor learning. Using diffusion MRI to infer structural plasticity, we showed that both the classic VMA paradigm—involving explicit and implicit processing—and an implicit-only version, engaged the left posterior hippocampus. Notably, while the classic paradigm induced transient changes in microstructure, implicit adaptation led to persistent structural alterations. These distinct temporal dynamics of hippocampal plasticity were mirrored by key motor regions involved in adaptation pointing to a close interaction between memory systems. Collectively, our findings demonstrate an active role of the left hippocampus in the form of procedural motor learning engaged by the VMA paradigms.

In the last decade, functional imaging, neuropsychological studies, and more recently electrophysiological work have reported the participation of the human hippocampus in motor skill learning, particularly within the domain of MSL (Fletcher et al., 2005; Albouy et al., 2013; Döhring et al., 2017; Long et al., 2018; Schapiro et al., 2019; Jacobacci et al., 2020a; Deleglise et al., 2023; Chen et al., 2024; Sjøgård et al., 2024). Given that MSL initially involves consciously memorizing the sequence, it is possible that the participation of the hippocampus in motor learning relates to its canonical declarative role. Our results demonstrating the involvement of this structure in implicit VMA are against this possibility and instead point to a more general function of this structure in memory encoding. Our work aligns with contemporary perspectives from human and nonhuman studies positing a unifying role of the hippocampus in sequence generation, a fundamental feature inherent in most motor learning paradigms (Buzsáki and Tingley, 2018; Rueckemann et al., 2021; Schwartenbeck et al., 2023).

Notably, changes in microstructure induced by MSL described in our previous study (Jacobacci et al., 2020a), and VMA/IVMA reported herein, were consistently localized to the posterior portion of the left hippocampus. The left-lateralization pattern may reflect the asymmetric control of the motor system in planning skilled and reaching movements (De Renzi and Lucchelli, 1988; Schaefer et al., 2007; Merrick et al., 2022; Yang et al., 2024). This hypothesis aligns with the observation that both motor and limbic structures were modulated to a similar extent in both tasks. But how are the motor system and the posterior hippocampus connected? Anatomical evidence indicates that the anterior and posterior hippocampi form part of two distinct networks associated with processing the “what” and “where” aspects of stimuli, respectively (Rolls et al., 2023). Specifically, the posterior portion of the hippocampus is connected with the retrosplenial cortex and the precuneus, which in turn, connects with the PPC involved in motor planning and sensorimotor adaptation (Della-Maggiore et al., 2004; Kahn et al., 2008; Margulies et al., 2009; Vesia et al., 2010; Adnan et al., 2016; Dalton et al., 2022). Thus, the consistent topography of hippocampal structural changes observed across motor paradigms may result from the connectivity between limbic and motor systems. Future studies may integrate tract-level analyses to test whether changes in gray-matter microstructure are accompanied by plasticity in connecting white matter pathways.

Interestingly, changes in hippocampal MD detected shortly after training correlated with the rate of adaptation in both VMA and IVMA, suggesting that they may be driven by the implicit component. This result together with the differing MD dynamics observed across tasks once memory consolidates (>6 h post-learning; Lerner et al., 2020; Solano et al., 2024) can be interpreted in light of the competition model, which proposes that implicit and explicit learning processes compete for error correction during VMA (Albert et al., 2022; Tsay et al., 2022). According to this model, an increase in the explicit system's response would reduce the extent of implicit adaptation. While we did not assess this directly in VMA, the experimental manipulation implemented here likely led to a greater extent of implicit adaptation compared with the classic paradigm (Morehead et al., 2017). Overall, our results suggest that, in the context of VMA, lasting changes in MD within the network involving the left hippocampus and key motor regions are contingent on the relative dominance of the implicit learning component.

What might the differing temporal dynamics in MD reflect at the biological level? Similar to MSL (Jacobacci et al., 2020a), VMA induced a transient MD decrease, whereas IVMA led to structural changes that persisted overnight. Yet, it is important to keep in mind that DTI models the diffusion signal at the voxel level and, therefore, does not provide direct information on the biological substrates of neuroplasticity. In fact, since MD reflects a combination of intra- and extracellular water diffusion without anatomical specificity, its interpretation as a marker of structural plasticity remains indirect (Assaf et al., 2004; Jones et al., 2013; Thomas and Baker, 2013). Nevertheless, studies linking in vivo DTI metrics with ex vivo histological markers in rodents provide insight into its potential biological underpinnings, suggesting that macroscopic MD decreases are associated with astrocytic expansion and/or remodeling (Blumenfeld-Katzir et al., 2011; Sagi et al., 2012; Hofstetter and Assaf, 2017). This process may be driven by various biological phenomena, including acute changes in neuronal activity (Darquié et al., 2001; Le Bihan et al., 2006; Abe et al., 2017), synaptic potentiation (Sagi et al., 2012), or synapse regulation and stabilization (Kleim et al., 2007; Stevenson et al., 2021). While neuronal activity associated with task execution is unlikely to explain MD changes 30 min later, other functional phenomena such as neuronal and astrogial swelling induced during LTP could rather impact on cell structure impacting on MD (Le Bihan et al., 2006; Sagi et al., 2012; Jin et al., 2013; Abe et al., 2017; Mader and Brimberg, 2019). In other words, early changes in microstructure may indeed reflect functional plasticity. In contrast, the enduring changes induced by IVMA may reflect distinct processes involving fiber organization, such as increased astroglial complexity associated with the stabilization of potentiated and/or new synapses (Kleim et al., 2007; Stevenson et al., 2021).

While our findings rely mostly on gray-matter microstructure, several facts confirm the specificity and reliability of our work. First, rapid MD changes induced by MSL and VMA tasks were associated with the rate of learning, pointing to the hippocampal involvement in the encoding/early consolidation of sensorimotor memories. Second, all three tasks induced structural changes that converged to the posterior portion of the left hippocampus supporting a role of this subfield in motor learning. Third, the whole-brain analysis identified regions of the motor system that are critically involved in VMA such as the motor cortex, the PPC, and the cerebellum (Della-Maggiore et al., 2004; Tseng et al., 2007; Criscimagna-Hemminger et al., 2010). Note that while it would have been optimal to relate changes in microstructure to functional variations in the same regions, fMRIs acquired during VMA present critical limitations (Diedrichsen et al., 2005). Specifically, during initial practice in VMA, increases in activity result from a combination of factors including changes in movement kinematic, error-driven processes, and force that may mask and/or confound the interpretation of learning-related changes in hippocampal activity. Thus, we could only have acquired and analyzed data obtained during IVMA, in which these parameters remain mostly constant.

In conclusion, our study underscores the specific engagement of the hippocampus in implicit motor learning. The convergence of structural changes observed in both visuomotor and MSL paradigms points to a broader involvement of the hippocampus that extends beyond declarative learning. Moreover, all motor paradigms recruited the left posterior hippocampus in temporal conjunction with key motor regions, a finding in line with the left-lateralized control of movement planning. Finally, our results suggest that in VMA, persistent changes in microstructure depend on the reliance on the implicit learning component. Our work points to the existence of common mechanisms in memory encoding across different domains regardless of the explicit or implicit nature of the task.