Working memory (WM), the capacity to maintain and manipulate information in a temporary mental buffer, is central to many aspects of human cognition. Indeed, through the interface between long-term memories and moment-to-moment information available in the environment, WM allows humans to organize relevant information in order to carry out successful goal-directed behaviors [1
]. As such, WM capacity is intrinsic to many daily activities such as reading, performing arithmetic, and keeping track of ideas during a conversation [2
]. Both WM capacity and the ability to manipulate content that is held in WM declines with age [4
]. Therefore, different approaches have been proposed to prevent this decline.
Brain stimulation techniques, such as repetitive transcranial magnetic stimulation (rTMS) and transcranial alternating current stimulation (tACS), have gained increased attention as a means to enhance WM and slow age-related impairments. rTMS uses brief, high-intensity magnetic fields to depolarize neurons underneath a magnetic coil. When applied over a brain region that helps support a specific cognitive function, rTMS has the potential to modulate related behavior. In many such studies, rTMS applied online, i.e., during the performance of a task, has been shown to interfere with ongoing cognitive processes, thus impairing behavioral performance ([6
] for a review). In other cases, however, studies have reported performance enhancement when applying online rTMS ([7
] for a review). For example, stimulation of the parietal cortex during WM maintenance tasks has resulted in significant decreases in reaction times and improvements in response accuracy [8
]. These contrasting results suggest that online rTMS may affect performance in a manner that is specific to the ongoing process and the spatio-temporal parameters of stimulation, for example, by modulating endogenous task-related oscillatory dynamics [10
Theta oscillations, in particular, have been shown to play a crucial role in memory processes [11
], and studies applying brain stimulation within the theta band (5–8 Hz) tend to demonstrate significant WM performance enhancement, as reported in a recent review investigating the effect of different types of brain stimulation, including tACS and rTMS [12
]. For example, using tACS, a neuromodulatory technique that uses oscillatory sinusoidal electrical currents to entrain the brain at a chosen frequency, Violante et al. [13
] demonstrated that synchronous theta-tACS (6 Hz) applied over the right fronto-parietal network, while subjects were performing either a simple choice reaction time task, a verbal 1-back task, or a verbal 2-back task, significantly shortened subjects’ reaction times. This effect was found only in the most difficult of the three tasks, with behavioral improvements that correlated with increased BOLD activity in the stimulated network, suggesting that tACS can improve both behavior and alter brain activity in a state-dependent manner.
The effects of rTMS on WM have also been investigated and have demonstrated positive effects on behavioral performance. Indeed, by applying 5 Hz rTMS over the left intraparietal sulcus during an auditory WM task, combined with electroencephalography recordings, Albouy et al. [14
] demonstrated that rTMS significantly increased the amplitude of theta oscillations and that this increase correlated with behavioral performance improvements observed with active stimulation. A similar pattern of results was also found by Li et al. [15
], who reported that 5 Hz rTMS applied over the left superior parietal lobe during a verbal WM task significantly improved behavioral performance, and increased theta-oscillations, while also reducing the amplitude of alpha oscillations. Moreover, these oscillatory physiological changes in the theta oscillations were positively correlated with behavioral improvements, indicating a mechanistic link between theta band activity and memory function.
In addition to periodic stimulation protocols that utilize a single oscillatory frequency, recent protocols have begun to test patterned stimulation protocols that involve multiple, simultaneous oscillations at different frequencies. In particular, intermittent theta-burst stimulation (iTBS), which utilizes intermittent patterns of 50 Hz pulses modulated at 5 Hz, has been shown to induce longer-lasting therapeutic effects for major depressive disorder, with shorter durations of stimulation. Studies that have tested iTBS in the context of WM have produced mixed results, with studies showing both large and significant behavioral improvements [16
], as well as small, non-significant effects [17
]. Despite the widespread interest in the brain stimulation to enhance WM, multiple methods being tested, and the development of more physiologically informed approaches, overall effect sizes in these studies remain moderate, as reported in a recent meta-analysis [18
Beyond the technical challenges of modulating WM performance with non-invasive brain stimulation, the application of neuromodulation in older adults presents a number of unique challenges. For example, aging has been associated with a relative decrease in the excitability of intracortical inhibitory circuits [19
], and is also associated with a decline in cortical thickness [20
] and cortical volume [21
], pointing towards a need to adjust for these factors when stimulating older adults. While linear adjustments of stimulation amplitude according to the distance from scalp to cortex [22
] may provide some correction for the systemic differences between older and younger adult brains, they do not account for the spread of rTMS-induced electric field across the cortical surface. In fact, there is reliable evidence that once multilinear changes in neuroanatomy associated with aging are controlled, the motor-evoked response to TMS does not differ significantly across the lifespan [23
], suggesting that electric-field (E-field) modeling is a necessary component to precise dosing of TMS attempting to induce changes in individuals with differing cortical geometry.
Recently, our group tested the effect of 5 Hz rTMS over the left dorso-lateral prefrontal cortex (DLPFC) of healthy young and older adults while participants performed a delayed-response alphabetization task (DRAT), in which they were asked to mentally arrange an array of letters into alphabetical order during a delay period. Results from this study revealed that both younger and older adults showed enhanced accuracy on the DRAT with active rTMS compared to somatosensory-matched electrical sham stimulation. These gains, however, were observed only in the most difficult task conditions, therefore replicating the difficulty-specific effect observed by Violante et al. [13
], reported above. Moreover, despite the presence of behavioral gains, the observed effects were small, leading to only a 4% improvement in memory recall [24
As such, the current study aimed to implement a related stimulation protocol in older adults, using a modified targeting approach with the goal of obtaining larger behavioral gains. Given past reports of rTMS-induced performance improvements in WM with stimulation delivered to the parietal cortex (e.g., [9
]), and extant theories that this region is central online attentional processes critical for WM [25
], this region was targeted in the current study. Moreover, because results from our previous fMRI study indicated that manipulation of information in the DRAT produced the greatest activation in the superior parietal lobule [27
], fMRI-BOLD activity in this region was used to derive individualized targets for each participant in the current study. Furthermore, in an attempt to induce approximately the same field strength in the target region across subjects, the current study defined stimulation amplitude according to E-field calculations, rather than resting motor threshold, as is typical of many studies. This computational dosing method accounts for the individual head anatomy and was deployed in an effort to minimize individual variability of the response.
Based on this experimental design, it was hypothesized that active rTMS would significantly enhance WM manipulation performance and that this effect would be most pronounced in the most difficult task conditions, consistent with the view that cognitive performance is most vulnerable to neuromodulation under the most demanding conditions [9
]. Contrary to this a priori hypothesis, 5 Hz rTMS to the left lateral parietal cortex during the delay period of the DRAT was found to impair WM performance. When considered in light of our former study [24
], the effects can be interpreted as being site-specific effects of rTMS on WM manipulation.
2. Materials and Methods
Thirty-nine healthy subjects (60–80 years old) were recruited into this single-blind randomized within-subject controlled trial, which was pre-registered on ClinicalTrials.gov
(NCT02767323). All participants provided written informed consent, which was approved by the Duke University Institutional Review Board (#Pro00065334). Participants were excluded if they had any contraindication to TMS or MRI, current or past psychiatric disorders or neurological disease (n
= 1), or a total scaled score lower than eight on the Dementia Rating Scale-2 [30
= 1). Participants were also excluded if they tested positive on a urine drug screen (n
= 2), performed poorly on the WM task during the initial visit (n
= 11), or experienced non-compliance, for example by responding in the task with random keys presses (n
= 3) (See Figure 1
for consort diagram). All data were collected in Duke School of Medicine, in the Brain Stimulation Research Center. According to these criteria, 18 participants were excluded, along with another 6 participants who withdrew participation for unspecified reasons. Fifteen subjects completed the full protocol (see Table 1
for baseline demographic). Participants had normal, or corrected-to-normal, vision and were native English speakers. Participants were compensated $
20/hour for their efforts with a $
100 bonus if they completed all study activities.
2.2. Experimental Protocol
Participants were scheduled for 6 sessions: a consenting visit including screening, resting motor threshold assessment, and practice at the delayed-response alphabetization task, followed by an MRI visit and four rTMS visits (Figure 2
). The following sections give a brief overview of the methods, and additional information can be found in [24
2.3. Delayed-Response Alphabetization Task (DRAT)
On each trial of the DRAT, an array of letters was presented on the screen for 3 s, followed by a 5-s delay period during which the participants were asked to mentally reorganize the letters into alphabetical order (Figure 3
). After the delay period, a letter with a number above it appeared on the screen for 4 s and participants were asked to report via a key press whether the letter was (1) not in the original set, (2) in the original set and the number matched the serial position of the letter once the sequence was alphabetized, or if (3) in the original set but the number did not match the serial position of the letter once alphabetized. These conditions are referred to as ‘new’, ‘valid’, and ‘invalid’, respectively.
During the first visit, participants performed the DRAT using a staircase procedure to establish individualized difficulty levels. Four individualized difficulty levels were defined according to the intersection between a sigmoid curve, fitted to the data, and an 82% accuracy threshold. The two set sizes below this intersection were defined as ‘very easy’ and ‘easy’, and the two levels above it were defined as ‘medium’ and ‘hard’. If the intersection between the curve and the threshold was lower than four, participants were considered poor performers and excluded from the study (n = 11). While all four difficulty levels were used for the subsequent imaging visit, only the ‘easy’ and ‘hard’ levels were used during the TMS visits.
2.4. Targeting Approach
During the second visit, subjects participated in an MRI scanning (General Electric MRI scanner, B0 field strength = 3 Tesla) during which structural-T1-weighted (echo-planar sequence: voxel size = 1 mm3, TR = 7.148 ms, TE = 2.704 ms, flip angle = 12°, FOV = 256 mm2, bandwidth = 127.8 Hz/Pixel), T2-weighted (echo-planar sequence with fat saturation: voxel size = 0.9375 × 0.9375 × 2.0 mm3, TR = 4 s, TE = 77.23 ms, flip angle = 111°, FOV = 240 mm2, bandwidth = 129.1 Hz/Pixel), and diffusion-weighted scans (single-shot echo-planar: voxel size = 2 mm3, TR = 17 s, TE = 91.4 ms, flip angle = 90°, FOV = 256 mm2, bandwidth = 127.8 Hz/Pixel, matrix size = 1282, B-value = 2000 s/mm2, diffusion directions = 26) were obtained. Functional acquisitions (EPI-sequence: voxel size = 3.4375 × 3.4375 × 3.99 mm3, TR = 2 s, TE = 25 ms, flip angle = 90°, FOV = 220 mm2, bandwidth = 127.7 Hz/Pixel) were also acquired as participants performed the DRAT in the scanner. After preprocessing the images, functional data were analyzed using a general linear model (GLM) in which trial events were convolved with a double-gamma hemodynamic response function. The GLM examined BOLD responses during trials where the correct response was chosen in the behavioral task. Separate events were modeled for the array presentation (3-s duration), the delay period (5-s duration), and the response period (trial-specific RT duration). All incorrect and non-response trials were modeled identically, but separately, and were not considered in the results.
The stimulation target was individually defined as the peak activation within the left lateral parietal cortex associated with a parametric increase in task difficulty during the delay period of the DRAT. According to the results obtained in our previous study [27
], both set size (the number of letters in an array) and sorting steps (the minimum number of operations required to alphabetize the array) contributed to the difficulty of an individual trial. Therefore, to obtain a more accurate representation of increases in DRAT difficulty, a parametric delay-period regressor, defined by the interaction between set size and sorting steps, was used to estimate task difficulty. This parametric regressor was orthogonalized to the non-parametric delay-period regressor. At the first level, functional data were analyzed as individual runs. Second-level analyses combined data across runs for each subject using a fixed-effects model. This processing allowed for the definition of the stimulation target on individualized statistical maps that predicted the parametric increase in BOLD activity associated with increasing task difficulty if the peak activation reached a z
-statistic value >2; or alternatively on the nonparametric delay-period map, if the peak did not reach this significance threshold.
To constrain the stimulation target within the left lateral parietal cortex, a mask obtained from the group activation of 22 older adults who participated in our previous study was used [24
]. The mask was defined as the overlap between the parametric interaction between set size and sorting steps (at z
> 1) and the non-parametric delay period activity (at z
> 1), therefore reflecting cortical regions that were generally activated by the task, but also specific to difficulty increase. The individual activation was then transformed back into subject space, and the peak activation within this mask was selected as the TMS target in the neuronavigation system (BrainSight, Rogue Research, Canada).
To define the coil orientation, the coordinates from the TMS target were projected onto the scalp surface using a nearest neighbor approach and then projected slightly outwards to account for the subject’s hair thickness (Supplementary Figure S1
). Hair thickness was measured on each participant during the screening visit, using a depth gauge (Digital Tread Depth Gauge, Audew, Hong Kong; resolution 0.01 mm) installed on a custom-made plastic base placed over the center of the group parietal mask (Supplementary Figure S4
). The TMS coil was then oriented around the scalp normal vector so that the direction of the second phase of the induced E-field coincided with the inward-pointing normal vector on the sulcal wall closest to the brain target location. This pulse direction induced the strongest E-field and activation at the target [31
]. The sulcal wall was identified using Freesurfer’s gyral/sulcal cortex classification ([33
]: file lh.sulc), a byproduct of SimNIBS’ mri2mesh script during the brain surface extraction) and a brain surface point was chosen at the transition location in-between local concavity and convexity (|local curvature threshold| < 0.05) defining the sulcal wall. In order to compute the normal vector of that sulcal wall point, the surface normal of the triangles were averaged. The intended coil orientation was then entered in the neuronavigation system using the ‘twist’ tool.
2.5. Stimulation Amplitude Approach
Rather than defining rTMS pulse amplitude according to a percentage of the motor threshold, as is frequently done in the literature, amplitude here was defined according to target-specific E-field values. While the motor threshold provides individualized information regarding the cortical reactivity of the motor cortex, it does not take into account differences in head anatomy and brain physiology between the motor cortex and other cortical regions within an individual. As such, traditional amplitude calibration based on the motor threshold may lead to substantial variation in the desired E-field strength in the targeted brain region. This may lead to response variability since the E-field strength is the key determinant of neural activation by TMS [34
]. Therefore, in the present study, we selected the TMS pulse amplitude (coil current rate of change, di/dt) to induce a specific E-field magnitude, Eref, in the left lateral parietal region of interest (ROI) across subjects.
To define Eref, computer simulations were used to estimate the E-field distribution within the parietal ROI induced when TMS was applied at an amplitude equal to the resting motor threshold in each of 9 subjects from a previous study [24
] (see Supplementary Figure S2
). For each of the 9 subjects, a parietal ROI was constructed by individual fMRI activity (|z| > 0; within a group activity mask) registered to the individual’s space (FSL flirt [35
]) within voxels classified as gray matter (SimNIBS: gm_only.nii.gz). The selected voxels were ranked according to their E-field strength, and a metric, E100, was defined as the minimum strength across the 100 voxels with the strongest E-field (Supplementary Figure S3
). The meaning of this metric is that 100 voxels in the ROI, corresponding to a volume of 100 mm3
, have E-field strength larger than E100. The average E100 across the 9 subjects was calculated to be 56 V/m, which we set as our desired target E-field strength, Eref = 56 V/m.
To select the individual TMS pulse amplitude in this study, computer simulations were performed to estimate the individual E-field distribution (analyzed for the left parietal ROI) and determine a TMS coil di/dt for which E100 = Eref in the ROI for each subject. Since TMS pulse di/dt scales linearly with the induced E-field, TMS was simulated for di/dt = 106 A/s, and a scaling factor was computed for di/dt to reach Eref for the hair thickness measured during the screening visit. The individual’s di/dt-value was determined for different hair thicknesses (in steps of 0.5 mm from scalp surface) and stored in a reference table (Supplementary Table S1
). During the first TMS visit, the hair thickness at the exact stimulation location was re-measured and rounded to match the closest value in the table. The corresponding computed di/dt value was selected. The TMS amplitude, expressed as a percentage of the maximum stimulator output (MSO), was adjusted for the chosen di/dt value. The amplitude, together with the determined location and orientation (described in Section 2.4
), were then experimentally applied. Two E-field strengths in the targeted region were experimentally tested, with E100 metric equal to either 80% Eref or 100% Eref. Resting motor threshold assessed during the screening visit was used to ensure that all stimulation intensities were below 130% of the resting motor threshold, and therefore within the published safety guidelines [36
The computer simulations of the TMS-induced E-field were performed using the SimNIBS software package (Version 2.0.1; [32
]). A computational model of each participant’s head was first generated employing co-registered T1- and T2-weighted MRI data sets to model major head tissues (scalp, skull, cerebrospinal fluid, gray and white brain matter) represented as tetrahedral mesh elements. Each mesh element was assigned a conductivity value based on its tissue association. The scalp, skull, and cerebrospinal fluid conductivities were set to isotropic values of 0.465, 0.010, and 1.654 S/m, respectively. The gray and white matter compartments were assigned anisotropic conductivities to account for the fibered tissue structures. This was accomplished within SimNIBS by co-registering diffusion-weighted imaging data (available for 14 out of 15 participants) and employing a volume normalization approach [37
] which kept the geometric mean of the conductivity tensor eigenvalues equal to the default literature-based isotropic values of 0.275 and 0.126 S/m for gray and white matter, respectively. For the subject with missing DTI information, the latter values were assigned as isotropic conductivities. (Supplementary Table S2
for individual subjects’ information)
2.6. TMS Procedure
During visits 3 to 6, participants performed the DRAT while active or sham rTMS was delivered to the individualized left lateral parietal target using an A/P Cool-B65 coil (MagVenture, Alpharetta, GA, USA). Twenty-five pulses of 5 Hz rTMS were delivered during the delay period of each trial (Figure 3
). For every two trials with stimulation, one trial without stimulation was performed. This approach, successfully used in multiple studies by Luber et al. [9
], theoretically allows time for neural activity in the stimulated region to return to its homeostatic baseline, allowing for greater range for the production of rTMS-induced plasticity and thus, potentially, greater rTMS effect on behavioral performance. The non-stimulated trials were excluded from subsequent analyses. The two intensities of stimulation, 80% Eref and 100% Eref, were applied on different days. Sham stimulation was applied using the same coil in placebo mode, which produced clicking sounds and somatosensory sensation via electrical stimulation with scalp electrodes similar to the active mode, but without a significant E-field induced in the brain [40
]. This type of sham stimulation allows participants to stay blinded during the course of the experiment. Neuronavigation (BrainSight, Rogue Research, Canada) and real-time robotic control (SmartMove, ANT, the Netherlands) were used to ensure that the optimal coil position was maintained throughout the stimulation sessions.
On each TMS visit, subjects performed the DRAT at their two individually titrated difficulty levels (‘easy’ and ‘hard’). Twelve blocks of the DRAT task were performed (30 trials per block): one block without stimulation (No-Stim1), followed by five blocks of active or sham stimulation, one block without stimulation (No-Stim2), and five more blocks with the sham or active stimulation. The first 5 rTMS blocks in the first visit were always active rTMS at 100% Eref to ensure that the subjects were able to tolerate this stimulation amplitude, with the later 5 rTMS blocks being sham stimulation at output setting corresponding to the 100% Eref condition. For the three other visits, the intensities of stimulation were alternated by day, and sham and active stimulation were applied on the same day in random order. Random allocation, enrollment, and assignment was made through a Matlab script that was administered by the clinical research specialists. No adverse events or pain due to the stimulation were reported by any of the subjects. As noted above, our central hypothesis was that older adults would show a benefit for WM accuracy on the DRAT due to online rTMS, but only during the most difficult condition.
2.7. Statistical Analyses
Analyses were performed using the general linear model module of Statistica (TIBCO Software Inc., Palo Alto, CA, USA), normality was tested using Kolmogorov–Smirnoff tests, and multiple comparisons corrections were performed using Bonferroni correction. All the results are expressed as mean ± standard error.