Abstract
The speed-accuracy trade-off (SAT), whereby faster decisions increase the likelihood of an error, reflects a cognitive strategy humans must engage in during the performance of almost all daily tasks. To date, computational modeling has implicated the latent decision variable of response caution (thresholds), the amount of evidence required for a decision to be made, in the SAT. Previous imaging has associated frontal regions, notably the left prefrontal cortex and the presupplementary motor area (pre-SMA), with the setting of such caution levels. In addition, causal brain stimulation studies, using transcranial direct current stimulation (tDCS), have indicated that while both of these regions are involved in the SAT, their role appears to be dissociable. tDCS efficacy to impact decision-making processes has previously been linked with neurochemical concentrations and cortical thickness of stimulated regions. However, to date, it is unknown whether these neurophysiological measures predict individual differences in the SAT, and brain stimulation effects on the SAT. Using ultra-high field (7T) imaging, here we report that instruction-based adjustments in caution are associated with both neurochemical excitability (the balance between GABA+ and glutamate) and cortical thickness across a range of frontal regions in both sexes. In addition, cortical thickness, but not neurochemical concentrations, was associated with the efficacy of left prefrontal and superior medial frontal cortex (SMFC) stimulation to modulate performance. Overall, our findings elucidate key neurophysiological predictors, frontal neural excitation, of individual differences in latent psychological processes and the efficacy of stimulation to modulate these.
SIGNIFICANCE STATEMENT The speed-accuracy trade-off (SAT), faster decisions increase the likelihood of an error, reflects a cognitive strategy humans must engage in during most daily tasks. The SAT is often investigated by explicitly instructing participants to prioritize speed or accuracy when responding to stimuli. Using ultra-high field (7T) magnetic resonance imaging (MRI), we found that individual differences in the extent to which participants adjust their decision strategies with instruction related to neurochemical excitability (ratio of GABA+ to glutamate) and cortical thickness in the frontal cortex. Moreover, brain stimulation to the left prefrontal cortex and the superior medial frontal cortex (SMFC) modulated performance, with the efficacy specifically related to cortical thickness. This work sheds new light on the neurophysiological basis of decision strategies and brain stimulation.
Introduction
The faster we make decisions, the more likely we are to make an error. This ubiquitous phenomenon, known as the speed-accuracy trade-off (SAT), reflects a decision strategy humans engage in during the performance of almost all daily tasks, from the simple (e.g., cooking, driving) to the critical (e.g., surgery, military command). In some contexts, it may be more appropriate to prioritize accuracy, for example, when driving down a busy road, but under other circumstances, speed may have greater importance, for example, if you have a passenger experiencing a medical emergency (and you must get them to a hospital). The cognitive and neural underpinnings of this simple strategic process have been the topic of intense scientific investigation (Heitz, 2014) and the generalizability of the SAT has even been demonstrated in climbing plants as they assess the strength of the supports they coil around (Ceccarini et al., 2020).
To assess the SAT, research has predominantly used simple decision-making paradigms (e.g., discriminating the motion direction of a cloud of dots) with varying task instructions: prioritize speed, prioritize accuracy, or balance the two. Computational modeling [e.g., the linear ballistic accumulator (LBA) model; Brown and Heathcote, 2008] has been applied to assess underlying latent process(es) involved in the SAT. Specifically, reaction time and accuracy data can be decomposed into latent decision variables of drift rates (evidence accumulation rates) and response caution (threshold; see Fig. 1). The SAT has predominantly been associated with changes in response caution (Palmer et al., 2005), whereby speed focused decisions have lower caution than accuracy focused conditions, although drift rates may also play a role (Sewell and Stallman, 2020).
Schematic of the linear ballistic accumulator model.
In terms of the underlying neurophysiology associated with the SAT, imaging studies have implicated activity in, and the structure of, frontal cortices. Specifically, task instruction manipulations of the SAT have been linked to activity in the presupplementary motor area (pre-SMA) and striatum (Forstmann et al., 2008), and left prefrontal cortex (Ivanoff et al., 2008; Vallesi et al., 2012). Studies with patients who have brain tumors have also implicated the left prefrontal (and not right) in the setting of response caution (Campanella et al., 2016). Individual differences analyses have correlated the strength of structural (white matter) connectivity between the pre-SMA and the striatum to the degree of adjustment to response caution with instruction (Forstmann et al., 2010). This work has been instrumental in developing our understanding of the SAT; however, imaging methodology is inherently correlational and previous studies have generally used small sample sizes (e.g., n = 9; Forstmann et al., 2010).
Taking a causal approach, we have previously shown that offline cathodal transcranial direct current stimulation (tDCS) applied to the left prefrontal cortex and the superior medial frontal cortex (SMFC; which includes the pre-SMA) can modulate decision strategies (Filmer et al., 2021): stimulating these two regions modulates response caution, but in opposing directions thus implying different processes/mechanisms of action. However, the efficacy of tDCS to modulate performance varies substantively between individuals (Wiethoff et al., 2014). This variability has been related to individual differences in underlying neurophysiology. Specifically, cortical thickness in regions proximal to the targeting electrode (Filmer et al., 2019b), and concentrations of neurochemicals, the ratio of the key inhibitory (GABA) and excitatory [glutamate (Glu)] neurochemicals, have been linked to frontal (Filmer et al., 2019a) and motor cortex (Stagg et al., 2011) stimulation outcomes.
To date, it is unknown what neurophysiological factors relate to individual differences in decision strategies, or stimulation efficacy to modulate these. We assessed whether the relative concentrations of GABA+ and Glu, and cortical thickness in frontal regions, related to decision strategies and modulations to the SAT via stimulation. Using ultra-high field magnetic resonance imaging (MRI; 7T), magnetic resonance spectroscopy (MRS) was acquired to estimate neurochemical concentrations in the left prefrontal cortex (and visual cortex as a control region), and T1-weighted (T1w) scans to estimate cortical thickness. To preview the results, decision strategy adjustment related to cortical thickness across a range of frontal areas and the ratio of GABA+ to Glu in left prefrontal cortex. Moreover, the efficacy of left prefrontal and SMFC stimulation to modulate decision strategies related to cortical thickness.
Materials and Methods
Experimental design
The experiment was based on an approach previously employed by Filmer et al. (2021). Participants who had previously undergone neuroimaging at 7T were recruited, with all subsequent testing preregistered (https://osf.io/b47dm/). Each participant completed three counterbalanced stimulation sessions in which they received a different form of cathodal stimulation: stimulation to the SMFC, stimulation to the left prefrontal cortex, and a sham condition. The task involved participants discriminating motion direction (left or right) from a random dot motion (RDM) stimulus, where they were instructed, across blocks, to prioritize speed, accuracy, or balance the two. Computational modeling (via the LBA; Brown and Heathcote, 2008) was applied to estimate latent decision-making variables (drift rates and response caution).
Participants
As per our preregistration, we aimed to recruit participants until one of the following occurred: we obtained data from 78 participants, our stopping date was reached, or we exhausted our participant pool. The latter two both occurred, with 38 participants tested in total (14 male, 24 female, mean age = 24, SD = 4.85). The subjects were screened in advance to ensure they could safely receive tDCS. In addition, any subjects with performance close to floor (<60%) or ceiling (>90%) for any of the three strategy conditions were excluded from the analysis. As a slight deviation from our preregistration, this was done on a session basis, so if performance met exclusion criteria for one session, only that session's data were removed. In the case that performance in the sham condition was affected, all data were removed as the sham session acted as baseline. This approach was taken to maximize data inclusion. In total, data were completely removed for two participants (one for performance <60%, one for performance >90%), active prefrontal data were removed for one further subject (performance <60%), and active SMFC data were removed for four subjects (one for performance <60%, three for performance >90%).
Task
A random dot motion (RDM) task with varying speed-accuracy condition instructions was employed. In each trial, participants viewed a display with a fixation dot (1000 ms), followed by the random dot display containing 200 dots (500 ms; see Fig. 2). The dot display was presented in a circular area with a diameter of 3.4° visual angle and the dots moved at a rate of 0.46° visual angle/s. The display contained a portion of dots (75%) moving coherently either to the top left or right of the screen, while the rest moved randomly. In each trial, subjects discriminated the direction of coherent motion. Each session started with a short practice block, followed by a thresholding procedure. The thresholding procedure used a QUEST staircase (Watson and Pelli, 1983) to achieve ∼70% accuracy via varying the degree of offset from the vertex for the motion. Three overlapping staircases were used, running for a total of 90 trials. A short check then occurred (20 trials), and if performance was <60% or >90%, the staircase procedure was re-run. Directly after thresholding, the subjects completed a short practice of the main task with one short (10-trial) block for each instruction condition: emphasize accuracy, emphasize speed, or balance the two.
Example trial. After an initial fixation of 1000 ms, a random dot motion (RDM) display was presented for 500 ms, then a fixation screen for a further 1300 ms. The color of the fixation dots indicated response instruction: red to focus on accuracy focus, green to speed focus, and orange to balance the two.
Next, subjects received stimulation (see details below). The main task then started within 2 min of the cessation of stimulation. This phase of the experiment consisted of 36 blocks (12 per instruction condition). The block types were pseudo-randomly intermixed, controlling for the direction and distance of instruction shifts, and ensuring the different block types occurred equally often across every 6 blocks of the experiment. The length of each block averaged 25 trials (minimum of 20, maximum of 30, with a SD of 3; ∼300 trials per response strategy per session). A 30-s break was given every six blocks. The total duration of the main task was 35 min.
tDCS
The tDCS procedure was identical to Filmer et al. (2021), wherein two different brain regions were targeted using the 10–20 EEG system: the SMFC (1 cm posterior to the Fz electrode) and the left PFC (1 cm posterior to the F3 electrode). In all sessions, the target electrode was the cathode, and the reference/return (anode) electrode was placed over the right orbitofrontal area (just above the eyebrow; Fig. 3). The electrodes were 5 × 5 cm saline-soaked sponges, with stimulation administered at 0.7 mA, a stimulation density that is common across the literature (Nitsche and Paulus, 2000; Filmer et al., 2013, 2021; Stagg et al., 2013).
Subjects received three different types of stimulation across the sessions: active stimulation to the left prefrontal cortex, active stimulation to the SMFC, or sham (placebo) stimulation. The location of the electrodes for the sham condition was controlled across subjects. Specifically, odd subject numbers received sham with the prefrontal electrode montage, and even subject numbers the SMFC electrode montage. The order of the sessions was counterbalanced across subjects. Subjects were blinded for all sessions. Because of the design, the experimenters were blinded for the sham session and one of the active stimulation sessions. That is, for those subjects receiving sham stimulation with the prefrontal montage, the experimenter was blinded for the active prefrontal and sham conditions, but not for the SMFC condition. We probed subjects, at the end of the final session, to ascertain whether they could discriminate the stimulation types.
For active stimulation, the current ramped up for 30 s, held constant for 19 min, then ramped down for 30 s. For the sham stimulation, the current ramped down after 15 s and administered small testing pulses to provide an online measure of impendence for the experimenter, continuing sensation for the participant, and thus give the illusion that the stimulation continues to run.
MRI
Across three imaging sessions, MR images were acquired on a 7T whole-body research scanner (Siemens Healthcare), with maximum gradient strength of 70 mT/m and a slew rate of 200 mT/m/s and a 7T Tx/32 channel Rx head array (Nova Medical). A total of 178 participants were scanned three times (in separate sessions) using an anatomic whole-brain T1w scan using a prototype MP2RAGE sequence (WIP 900; Marques et al., 2010; O'Brien et al., 2014) at 0.75-mm isotropic voxel size (TR/TE/TIs = 4300/2.5/840 ms, 2370 ms, TA = 6:54) as part of another project. A total of three participants did not have usable data (likely because of movement) and were excluded from the cortical thickness analyses.
T1w images were segmented using Advanced Normalization Tools (ANTs) version >=2.2.3. We first constructed a population-specific template (of 50 participants) using antsMultivariateTemplateConstruction2.sh and 20 iterations to achieve good convergence. The template was then labeled using the antsCookTemplatePriors script to construct our own tissue priors for all subsequent processing. Following this atlas construction, we used the antsLongitudinalCorticalThickness.sh script with default parameters for cortical thickness and white and gray matter labeling of each subject at each time point. Finally, we employed joint label fusion (JLF) with a subset of the Mindboggle-101 (Klein and Tourville, 2012; Klein et al., 2017) label priors to label each participants' DKT-31 regions (Klein and Tourville, 2012) at each time point. The ROIs included in this study were frontal regions across the two hemispheres (Fig. 4) as we anticipated they could be influenced by stimulation and involved in the SAT. Cortical thickness estimates and volume for these regions of interest (ROIs) were calculated using ANTs including mean, min, max, and SD cortical thickness values, surface area, and volume for the given ROIs. Our analyses focused on cortical thickness, and visual inspection of the cortical thickness images were completed. As each participant was scanned three times, we took the median cortical thickness values for each participant. The final data were checked for outliers (values >3 SD from the mean), and none were found.
MRS
Using the same scanner as above, anatomic T1-weighted images were acquired to allow for MRS voxel placement (MP2-RAGE sequence: TR = 4300 ms, TE = 3.38 ms). One voxel was placed over the left prefrontal cortex (40 × 26 × 26 mm), positioned on a slice 1.5 mm above the superior margin of the lateral ventricles with the center one third of the total anterior/posterior distance and halfway between the mid-line and the left lateral border (Fig. 5). The other voxel was positioned centrally across the bilateral midline, above and aligned to the calcarine fissure. To detect GABA+, a MEGA-sLASER sequence (Andreychenko et al., 2012) was used [key parameters: TE = 74 ms, TR = 7800 ms, bandwidth = 4000 Hz, 64 transients of 4096 data points, two water-unsuppressed transients also acquired, and a 23.6-ms Gaussian editing pulse was applied at 1.90 (ON) and 7.46 (OFF) ppm]. Water suppression occurred via variable power with optimised relaxation delays (VAPOUR, Tkáč and Gruetter, 2005). To detect Glu, a sLASER sequence (Scheenen et al., 2008) was used (key parameters: TE = 42ms, TR = 7790 ms, spectral width = 4000 Hz, 32 transients of 4096 data points acquired). In addition, a single inversion recovery metabolite nulling in sLASER (TI = 1135 ms) was performed in two subjects to estimate, and in the analyses constrain, metabolite quantification (Cudalbu et al., 2021). In quick succession, automated shimming with 3D GRE brain followed by FAST(EST)MAP shimming were run. The acquisition window for MRS sequences was 1000 ms; however, we only used the first 250 ms of the data for each acquisition as there was no signal in the free induction decay after this time. All MRS data were analyzed from the Siemens TWIX format.
We used the same procedures for spectral preprocessing and quantification as Rideaux et al. (2022). Briefly, all edited (MEGA) subspectra were 3-Hz exponentially line-broadened and zero-filled to a spectral resolution of 0.061 Hz/point. For all data, frequency and phase parameter estimates were obtained by modeling the total creatine (creatine and phosphocreatine) signal, then these parameters were used to align subspectra to a common frequency and phase, as previous work has shown this is the most effective method of alignment for GABA+ quantification (Rideaux et al., 2021). For edited data, the average ON and OFF spectra were subtracted to produce the edited spectrum, from which GABA+ (3 ppm) signal intensity was modeled using a Lorentzian. Similarly, water signal intensity was determined by fitting a Lorentzian model to the average water-unsuppressed spectra at 4.7 ppm. For unedited data, glutamate and water signal intensity were quantified using LCModel. GABA and glutamate concentrations were calculated as the area and the peak of the fitted peak, respectively, and expressed in institutional units (i.u.) using the unsuppressed water signal as an internal concentration reference. We further applied a tissue-correction method that accounts for differences in relaxation times across the tissue types within a voxel (Gasparovic et al., 2006).
It should be noted that the MRI/MRS data were collected for a separate project, and there was approximately one year in between imaging and the SAT/tDCS sessions. MRI/MRS measures have been shown to be stable over time (Near et al., 2014; Greenhouse et al., 2016); thus, our MRI/MRS measures remain robust and valid.
Statistical analyses
Computational modeling
Drift rates and response thresholds were estimated using the linear ballistic accumulator (LBA) model within a hierarchical Bayesian framework, which captures individual differences in process components while also providing information about each quantity at the group level. The model was identical to the one specified by Filmer et al. (2021) except that the mean drift rate for the incorrect response alternative was fixed across experimental conditions. This was done to enable a more reliable measure of discriminability and caution (see Evans, 2020).
Discriminability was indexed as the mean drift rate for the correct response alternative. Given that the mean drift rate for the incorrect response was fixed across conditions, changes in mean drift rate for the correct response across conditions also reflected changes in the difference in the mean drift rates between the correct and incorrect response alternatives. Response caution was indexed as the difference between the response threshold and the maximum starting point of evidence accumulation. Discriminability and response caution were allowed to vary across experimental conditions, while nondecision time, the maximum start point of evidence, and the sum of the correct and error drift rates were fixed across conditions. To ensure model identifiability, the SD of both drift rates was fixed to one in all conditions. We used 95% credible intervals to make inferences about group-level effects.
Individual differences analyses
Regression analyses were conducted to assess for associations between individual differences in the latent decision process components of the SAT and in the efficacy of tDCS to modulate these. These analyses utilised response caution and discriminability measures derived from the LBA model. The following variables were calculated. (1) The change in caution with instruction. (2) The change in discriminability with instruction. (3) Overall caution. (4) Overall discriminability. The two change measures were calculated by subtracting the value in the speed-focused condition from the value in the accuracy-focused condition, and the two overall conditions were the average across instruction. These variables were calculated for each session, and then active stimulation was compared with sham (e.g., the change in caution was subtracted for active vs sham sessions to derive the extent to which stimulation increased the change in caution) and used as the dependent variables in the analyses reported below. Independent variables were the concentrations of neurochemicals, and cortical thickness for frontal regions. For cortical thickness, we had a relatively large number of brain regions across the frontal cortex (a total of 12 regions; see Fig. 4) which made initial Bayesian regressions computationally intensive and thus impractical to run. Thus, we initially conducted linear backward regressions as a form of variable selection (stepping method criteria based on p values, entry of 0.05 and removal 0.1). The variables identified from this were then included in a Bayesian linear regression. All regression analyses were completed in JASP (JASP Team, 2017), and included order of session completion in the null model to account for the order of session completion. The regression analyses with cortical thickness also included total brain volume in the null model. The Bayes factors (BFs) were interpreted as follows (either in favor of the alternate, BF10, or null, BF01, hypotheses): 1–1.3 as indeterminant, 1.3–3 as weak evidence, 3–10 as moderate evidence, and >10 as strong evidence.
Results
Group level effects
Manipulation check
From the sham sessions, we found the instruction manipulation successfully modulated both response caution and discriminability (Fig. 6). Specifically, accuracy-focused instruction resulted in more cautious responses than mixed [0.098 0.151] or speeded [0.405 0.456], and mixed was also more cautious than speeded [0.283 0.33]. Speeded instruction resulted in lower discriminability than mixed [0.027 0.123] or accuracy [0.022 0.12], and mixed and accuracy were not credibly different [−0.052 0.043].
Stimulation effects on response caution
Overall, we replicated our previous findings (Filmer et al., 2021) showing that stimulation to the SMFC reduces response caution relative to sham [0.148 0.239] (Fig. 7). However, this effect did not interact with instruction (CIs for differences between instruction conditions all contained 0). We had anticipated that stimulation to the prefrontal cortex would increase caution, but in fact found the opposite [0.072 0.147], although the decrease was to a lesser extent than stimulation to the SMFC [0.035 0.134]. The effect of prefrontal stimulation interacted with instruction, such that the effect of stimulation was greater for accuracy compared with speeded instruction [0.01 0.157], with no other comparisons credible (all CIs contained 0).
Stimulation effects on discriminability
Our preregistered hypotheses focused on stimulation modulating caution. Nonetheless, we had previously found (and thus anticipated) that stimulation to the SMFC increased discriminability. Here, however, stimulating the SMFC did not credibly modulate discriminability compared with sham [−0.086 0.017], and there were no interactions between the instruction conditions (all intervals contained 0). Stimulating the prefrontal cortex modulated discriminability relative to sham stimulation [0.019 0.102], although there was no difference between prefrontal and SMFC sessions [−0.084 0.031]. For prefrontal stimulation (relative to sham) there were no differences between instruction conditions (all intervals contained 0). As the two active stimulation conditions were not distinct, overall, there is little evidence to support stimulation modulating discriminability.
Neurophysiological correlates of the SAT
Neurochemicals
For the sham stimulation sessions, the amount of adjustment in thresholds with instructions correlated with neurochemical concentrations (Fig. 8). Specifically, adjustment in threshold with task instruction related to E/I balance in the left prefrontal cortex (BF10 = 4.932, R2 = 0.2) with lower E/I ratio (more excitatory neurochemical balance) associated with greater response caution adjustment. No relationships were evident with neurochemical concentrations in the visual cortex (BF10 < 0.6 for all). Average caution, which could be thought of as how cautious overall an individual was in their response, was not related to E/I in either region of interest (PFC, BF10 < 0.56 for all; VC, BF10 < 0.9 for all). E/I was also not related to adjustment in discriminability for either region of interest (BF10 < 0.7 for all). There was only weak evidence for a relationship between mean discriminability and glutamate in the left prefrontal cortex (BF10 = 1.99, R2 = 0.132), and no evidence for any relationship for visual cortex (BF10 < 0.93 for all). Thus, overall, E/I balance was predominantly related to adjustments in thresholds with task instruction.
Cortical thickness
Overall, baseline cortical thickness in left and right frontal regions related to both caution and discriminability. Specifically, there was moderate evidence that the amount of adjustment in caution related to left and right rostral middle frontal, and left caudal middle frontal, gyri (BF10 = 4.631, R2 = 0.427; see Fig. 9) with thicker cortex in the left hemisphere leading to less, and the right hemisphere more, adjustment in caution. There was also strong evidence that the average caution related to right rostral middle frontal and left pars triangularis gyri (BF10 = 14.411, R2 = 0.419), with thicker cortex in the left pars triangularis, and thinner cortex in the right rostral middle frontal gyrus, relating to less overall caution. Thus, there is a pattern for an opposing relationship between left and right hemisphere cortical thickness and adjustments in, and overall level of, thresholds. Moreover, the middle frontal gyrus (in both hemispheres) appears to be an important region relating to response caution. This is in line with the previous literature: the middle frontal gyrus has been implicated in previous imaging studies (Ivanoff et al., 2008) and was a key target in previous tDCS studies (Filmer et al., 2021, 2023). But, further, it appears that there is some differing role for the two hemispheres.
For discriminability, there was moderate evidence that the adjustment with instruction related to a broad range of frontal regions across both hemispheres. Specifically, more adjustment in discriminability was associated with thicker cortex in right pars opercularis, left precentral and right rostral middle frontal gyri, and thinner cortex in left rostral middle frontal and right precentral gyri (BF10 = 5.299, R2 = 0.48). There was only weak evidence for average discriminability relating to cortical thickness (the left caudal middle frontal gyrus; BF10 = 1.52, R2 = 0.345).
Neurophysiological correlates of prefrontal stimulation efficacy
Neurochemicals
There was little support for neurochemical concentrations relating to the effect of prefrontal stimulation on thresholds or drift rates. There was indeterminate/weak evidence relating stimulation efficacy to modulate the adjustment of thresholds to E/I (BF10 = 1.384) and relating stimulation efficacy to modulate average discriminability to glutamate (BF10 = 1.428), and all other analyses were indeterminate/in favor of the null (BF10 < 0.5 for all).
Cortical thickness
The efficacy of stimulation to the prefrontal cortex (relative to sham) to modulate adjustments in discriminability and the overall level of caution were strongly related to cortical thickness across a range of frontal regions (Table 1). The regions relating to stimulation effects on discriminability and caution were distinct, implying some difference in sources of variability in the efficacy of stimulation to modulate these two processes. Overall, the regions relating to stimulation efficacy included both areas likely to be affected by stimulation (see the current modeling in Fig. 3) such as the left rostral and superior frontal gyri, and areas that show relatively less overlap with induced current from stimulation (e.g., right precentral), implying potential direct and indirect relationships between cortical thickness and stimulation outcomes. One possibility is that the areas not likely to be directly affected by stimulation could influence outcomes because of network-level effects of the stimulation that spread beyond target regions.
Winning regression models relating modulations to caution and discriminability with stimulation (relative to sham) with cortical thickness in frontal regions
Electrode montages and current models for the left prefrontal (A) and SMFC (B) target regions.
Key cortical thickness regions of interest for each of the stimulation conditions: superior frontal (yellow), rostral middle frontal (orange), caudal middle frontal (red), pars triangularis (green), pars opercularis (blue), and precentral (purple) gyri.
Locations of the MRS voxels in the frontal and visual cortex (A) and example spectra from the two MRS scan sequences (B).
Group level effects of instruction on response Caution (top) and Discriminability (bottom) for the sham stimulation session.
Group level effects of stimulation to prefrontal cortex (yellow) and SMFC (green) on caution (A, C) and discriminability (B, D) averaged across instruction condition (top) and brown down by instruction condition (bottom).
The ratio of GABA+ to glutamate in the left prefrontal cortex (A) related to how much participants adjusted their response caution with instruction. This was not the case for GABA+/glutamate in the visual cortex (B).
Cortical thickness in left (A) and right (B) rostral middle frontal gyri, and (C) left caudal middle frontal gyrus related to how much participants adjusted their response caution with instruction.
Neurophysiological correlates of SMFC stimulation efficacy
Neurochemicals
There was no evidence in favor of a relationship between the efficacy of stimulation to SMFC to modulate thresholds or discriminability (in terms of adjustments with instruction, or averages) and neurochemical concentrations (all BF10 < 1.05).
Cortical thickness
Modulations in the adjustment of caution and discriminability, and in overall level of caution, with stimulation (relative to sham) were associated with a broad range of frontal regions (see Table 1). There was substantial overlap in these regions between caution and discriminability effects, implying the source of variability for SMFC stimulation efficacy was consistent across latent decision variables. Again, the regions relating to stimulation efficacy included both areas likely to be affected by stimulation (see the current modeling in Fig. 3), such as right pars opercularis, pars triangularis, rostral middle frontal, and left caudal middle frontal gyri, and areas that do not largely overlap with induced current from stimulation (e.g., left pars opercularis, precentral, and rostral middle frontal gyri).
Blinding efficacy
At the end of the final session, participants were informed one of the three tDCS sessions involved a sham stimulation. We then asked participants to guess which session they received sham. Overall, there was a correct guess rate of 42%, which showed moderate evidence in favor of the null (BF01 = 3.2), i.e., participants could not correctly distinguish active from sham stimulation.
Discussion
We assessed neurophysiological factors that predict individual differences in decision strategies, and the efficacy of tDCS in modulating these factors. Using a behavioral paradigm where strategy was adjusted via task instruction (emphasize speed, accuracy, or balance the two), we partially replicated previous studies from our lab (Filmer et al., 2021, 2023). Specifically, we again found that stimulation to the SMFC with cathodal tDCS decreased the latent decision process of response caution relative to both sham and left prefrontal stimulation. However, we did not find that stimulation to the left prefrontal increased caution as we anticipated. Of import, performance at the SAT task was related to cortical thickness and neurochemical concentrations in the absence of active stimulation. Furthermore, cortical thickness but not neurochemical concentration was related to the efficacy of tDCS to modulate SAT performance, for both prefrontal and SMFC stimulation.
The study included two separate active stimulation conditions (in addition to sham stimulation), meaning it is unlikely the findings were because of something generic and/or driven by the location of the reference/return electrode. The design was also partially double blind (one active vs sham double blind for each participant), with participants unable to reliably guess which of the three sessions was sham. The sample size was smaller than the initial design (consisting of 34 participants), and there have recently been concerns raised about using relatively small sample sizes in brain-behavior association studies where small effect sizes are likely (Marek et al., 2022). However, these criticisms predominantly relate to studies using broader/more general measures of personality, psychopathology, behavior, and cognition (e.g., personality traits assessed via pen and paper tests; NIH Toolbox assessing general cognitive ability). Here, the research question was focused on a highly studied and specific cognitive phenomenon, the SAT, and each participant contributed a large number of trials (∼2700) to the analyses, resulting in individual participant measures that were relatively precise (see Gratton et al., 2022). Further, the data were analyzed using computational modeling to parse performance into latent decision variables. We have also coupled these more precise measures with a brain stimulation intervention, the use of Bayesian statistics, and ultra-high field (7T) neuroimaging giving high resolution for cortical thickness and spectral measures of neurochemical estimates. Thus, collectively, the study was powered to detect behavioral effect sizes and correlations typically seen in similar studies combining MRI, MRS and brain stimulation (Filmer et al., 2019b) and included the most precise behavioral and neuroimaging measures possible. Nevertheless, given the current questions regarding replicating brain-behavior associations, and that the true effect sizes are unknown, further research will be welcome to corroborate our findings.
The group level results did not fully replicate our previous findings, despite the results having been demonstrated twice before in other studies (Filmer et al., 2021, 2023). Specifically, we did not see the anticipated increase in response caution with tDCS to left prefrontal cortex. More broadly, the use of tDCS has been criticised because of reports of poor replicability (Horvath et al., 2014), and high interindividual and intraindividual differences in efficacy (Chew et al., 2015; Dyke et al., 2016; Wörsching et al., 2017). Indeed, the study reported here aimed to shed light on the individual differences in stimulation efficacy (also, see Filmer et al., 2020 for a discussion of the fields critique). It is worth noting that the study design was changed from our previous work. For example, the present study was changed to be fully within subjects, with each participant completing three (as opposed to two) experimental sessions with the task. This could have led to increased practice effects, weakening the findings. Also, we found substantial correlations between the effect of prefrontal stimulation and cortical thickness, showing there is a relationship here between the SAT and prefrontal stimulation. It is important to note that despite the lack of replication for the prefrontal cortex we did see a consistent effect from SMFC stimulation, for what is now the third time (Filmer et al., 2021, 2023).
Our data provide a novel insight into the SAT. We are the first to show that neurochemical E/I balance is related to adjustments in caution with instruction. Specifically, individuals who show more of an adjustment in caution with instruction had greater neurochemical excitation in the left prefrontal cortex. Adjustments in, and overall levels of, caution were also linked to cortical thickness in both left and right frontal regions. A clear pattern emerged where thicker cortex in the left, and thinner in right, was associated with less adjustment, and lower overall levels, of caution. A less clear pattern, involving a broader range of regions, was apparent for adjustments in discriminability.
Stimulation efficacy was not related to neurochemical concentrations, in contrast to findings considering stimulation effects on other performance measures such as response selection (Filmer et al., 2019a). It is possible the age of the MRS data reduced sensitivity for a possible relationship, although given previous findings relating to the stability of MRS (Greenhouse et al., 2016; Near et al., 2014), and the presence of a relationship with the SAT reported here, this seems unlikely. This implies that tDCS protocols to the frontal cortex can have differing underlying relationships with baseline neurophysiology depending on the targeted cognitive process. It is also possible that different neurochemicals could play a role in stimulation efficacy, for example, noradrenaline and dopamine, both not measurable with MRS, have been implicated in effects of tDCS on neuroplasticity (Nitsche et al., 2006; Kuo et al., 2017) and the efficacy of tDCS to modulate the SAT (Leow et al., 2023).
Strong relationships were evident, however, between cortical thickness and stimulation efficacy (see also Filmer et al., 2019b). Overall, a broad range of regions spanning both left and right hemispheres were found to be related. For SMFC stimulation, these regions largely overlapped between the different performance measures (adjustments in caution and discriminability, and overall level of caution). For left prefrontal stimulation, cortical regions relating to performance (adjustments in discriminability and average caution) showed less overlap, suggesting a potential difference in how these two effects were modulated. It is possible that stimulation efficacy relates to cortical thickness because of an influence on current flow. Thicker/thinner cortex could affect the dispersal of the current and thus the influence of stimulation on performance. Alternatively, it may be that cortical thickness relates to cognitive performance and the effect of stimulation on these underlying processes. These two are also not mutually exclusive, and could in combination explain the data reported here. Given that cortical thickness was found to relate to variability in SAT performance in the absence of stimulation (for the sham sessions only), it does appear that thickness is related to the SAT in some manner. Indeed, there was some overlap between the regions found to relate to adjustments in caution without stimulation, and regions relating to stimulation efficacy. In particular, the left caudal middle frontal gyrus was related to SAT performance in the absence of stimulation, and to tDCS efficacy. This is in line with both the current modeling showing induced electric fields in this region (Fig. 3), and also previous imaging work into the SAT (Ivanoff et al., 2008). Further work may look to distinguish between the possible contributions of cortical thickness to current dispersal and cognitive performance.
In summary, the work reported here gives unique insights into the neurophysiological variables relating to individual differences in SAT performance and stimulation efficacy. Given both the ubiquity of the SAT in everyday life, and the critique of tDCS for the individual variability in responsiveness (see Filmer et al., 2020), this work helps to elucidate the key mechanisms involved in both. As well as furthering understanding, the findings highlight how we may predict both an individual's SAT performance and how effective stimulation may be for modulating decision strategies. Indeed, in the future this could lead to individually tailored interventions to optimize performance.
Footnotes
This research was supported by Australian Research Council (ARC) Discovery Grants DP180101885 (to P.E.D.) and DP210101977 (to P.E.D. and H.L.F.). H.L.F. and T.B. were supported by ARC Discovery Early Career Researcher Awards DE190100299 (to H.L.F.) and DE180101340 (to T.B.).
The authors declare no competing financial interests.
- Correspondence should be addressed to Hannah L. Filmer at h.l.filmer{at}gmail.com
