|Home | About | Journals | Submit | Contact Us | Français|
Transcranial magnetic stimulation (TMS) to the left dorsolateral prefrontal cortex (DLPFC) is used clinically for the treatment of depression however outcomes vary greatly between patients. We have shown that average clinical efficacy of different left DLPFC TMS sites is related to intrinsic functional connectivity with remote regions including the subgenual cingulate and suggested that functional connectivity with these remote regions might be used to identify optimized left DLPFC targets for TMS. However it remains unclear if and how this connectivity-based targeting approach should be applied at the single-subject level to potentially individualize therapy to specific patients. In this article we show that individual differences in DLPFC connectivity are large, reproducible across sessions, and can be used to generate individualized DLPFC TMS targets that may prove clinically superior to those selected on the basis of group-average connectivity. Factors likely to improve individualized targeting including the use of seed maps and the focality of stimulation are investigated and discussed. The techniques presented here may be applicable to individualized targeting of focal brain stimulation across a range of diseases and stimulation modalities and can be experimentally tested in clinical trials.
Repetitive transcranial magnetic stimulation (rTMS) is a non-invasive brain stimulation technique that is showing utility in the treatment of a variety of neurological and psychiatric disorders (Burt et al., 2002; Fregni and Pascual-Leone, 2007; Hallett, 2007). Its most common use involves high-frequency stimulation to the left dorsal-lateral prefrontal cortex (DLPFC) for the treatment of depression (George et al., 1995; O'Reardon et al., 2007; Padberg and George, 2009; Pascual-Leone et al., 1996). In the US, the Neuronetics® device and Neurostar protocol are approved by the Food and Drug Administration for some patients with medication-resistant depression. However, the clinical utility of rTMS remains limited by large heterogeneity in clinical response.
One factor known to contribute to this response heterogeneity is differences in the site of stimulation in the left DLPFC (Ahdab et al., 2010; Fitzgerald et al., 2009; Herbsman et al., 2009; Herwig et al., 2001; Padberg and George, 2009). The targeting technique routinely employed in clinical practice, and followed by the FDA approved protocol, is to center the TMS coil at a point 5 cm anterior to the motor cortex measured along the curvature of the scalp. This approach identifies different stimulation sites in different subjects (Ahdab et al., 2010; Herwig et al., 2001) and some sites appear to be more effective than others at producing an antidepressant response (Fitzgerald et al., 2009; Herbsman et al., 2009; Padberg and George, 2009; Paillère Martinot et al., 2010). In an effort to understand why some sites are more effective, we recently used intrinsic (resting state) fMRI to identify differences in functional connectivity between effective and less effective DLPFC stimulation sites at the population level (Fox et al., 2012a). Significant differences in connectivity were seen in a variety of cortical and limbic regions including the subgenual cingulate, a region repeatedly implicated in antidepressant response to a variety of treatment modalities (Drevets et al., 2008; Mayberg, 2009; Mayberg et al., 2005). Specifically, left DLPFC sites which when targeted with TMS lead to greater antidepressant effects, showed a stronger negative correlation (anticorrelation) with the subgenual. Based on these findings, we proposed a connectivity-based targeting strategy for TMS and used this technique to identify theoretically optimal TMS target coordinates in the left DLPFC at the population level (Fox et al., 2012a).
An important advantage of this connectivity-based targeting strategy is that it might be scaled from the population level down to the level of single subjects to tailor treatment to individual patients. The DLPFC varies greatly between individuals on a histological basis (Rajkowska and Goldman-Rakic, 1995) thus the population-average TMS coordinates might be suboptimal for many patients. However individualized targeting will be associated with an inherent worsening of signal to noise that could overwhelm any benefit of accounting for individual differences. For example, targeting a population-average focus of hypometabolism in the left DLPFC with TMS appears superior to the standard 5 cm technique (Fitzgerald et al., 2009), however three separate studies aiming to target individualized foci of hypometabolism failed to provide clinical benefit (Garcia-Toro et al., 2006; Herwig et al., 2003a; Paillère Martinot et al., 2010). In fMRI, the subgenual is a region with poor signal to noise ratio (Ojemann et al., 1997) and intrinsic anticorrelations may be less reproducible than positive correlations (Shehzad et al., 2009). It therefore remains unclear if connectivity-based targeting can identify individualized TMS sites in the DLPFC that are sufficiently robust and reproducible to potentially be used in the treatment of depression.
In the current article we show that individual differences in DLPFC connectivity are large, reproducible across scanning sessions, and can be translated into individualized TMS targets on the cortical surface. Further, we identify factors likely to improve individualized targeting such as the use of seed maps and the focality of stimulation.
This study utilized three independent datasets, each of which was used for a particular set of analyses (Figure 1). Datasets 1 and 2 have been used in prior published articles (Fox et al., 2012a; Van Dijk et al., 2012; Yeo et al., 2011) and were not collected specifically for the present paper. However these prior articles investigated different experimental questions than those investigated in the current article and no part of the present report overlaps with prior published findings. Dataset 3 was collected specifically for the present investigation.
Two regions of interest (ROI) were defined a priori for use in the present analysis, one in the subgenual cingulate cortex and another in the left DLPFC. Details regarding the creation of these ROIs have been presented previously (Fox et al., 2012a). Briefly, the subgenual ROI was constructed by placing a 10 mm radius sphere at MNI coordinates (6, 16, −10), masked to include only sampled grey matter voxels. These subgenual coordinates were chosen based on prior studies where a reduction in subgenual activity was associated with antidepressant response across a wide range of treatment modalities (Drevets et al., 2002; Kito et al., 2008; Kito et al., 2011; Mayberg et al., 2000; Mayberg et al., 2005; Nahas et al., 2007; Wu et al., 1999) (see Table 1 in Fox et al 2012a).
A large ROI designed to cover what is generally considered the left DLPFC was constructed by centering 25 mm radius spheres on MNI coordinates at the center of Brodmann area (BA) 9 (−36, 39, 43), the center of BA46 (−44, 40, 29), and the average DLPFC site stimulated using the standard 5 cm targeting technique (−41, 16, 54). Spheres were summed and masked to include only sampled grey-matter voxels. Coordinates for Brodmann regions (BA9 and 46) were based on Rajkowska and Goldman-Rakic 1995 (Rajkowska and Goldman-Rakic, 1995). However this paper reported only y and z coordinates so the x coordinate was determined from the coordinate in Talairach space on the cortical surface constrained by the y and z coordinates (Talairach and Tournoux, 1988). This complete set of Talairach coordinates was transformed into MNI space using tal2mni (http://imaging.mrc-cbu.cam. ac.uk/imaging/MniTalairach). Coordinates for the average 5cm site were based on coordinates averaged from two prior TMS studies (Herbsman et al., 2009; Herwig et al., 2001).
fMRI data were processed in accordance with the strategy of Fox et al 2005 (Fox et al., 2005) as implemented in Van Dijk 2010 (Van Dijk et al., 2010). Briefly, the first four volumes of each run were discarded, slice-acquisition-dependent time shifts were compensated per volume using SPM2 (Wellcome Department of Cognitive Neurology, London, UK), and head motion was corrected using rigid body translation and rotation using FMRIB Software Library (Oxford, UK)(Jenkinson et al., 2002). Atlas registration was achieved by computing affine and nonlinear transforms connecting the first volume of the functional run using SPM2, with a T1 EPI template in the Montreal Neurological Institute (MNI) atlas space. Data were resampled to 2-mm isotropic voxels and spatially smoothed using a 6-mm full-width half-maximum (FWHM) Gaussian kernel. The data underwent further preprocessing specific to functional connectivity analysis, including low-pass temporal filtering (f < 0.08 Hz), head-motion regression, whole-brain signal regression, and ventricular and white matter signal regression using custom-designed in-house software (Fox et al., 2005; Van Dijk et al., 2010).
Three “seed regions” were used in the present functional connectivity analysis, all taken from a prior study of DLPF and subgenual functional connectivity (Figure 2) (Fox et al., 2012a). The first seed region was simply the small subgenual ROI as defined above (Subgenual ROI). The second seed was the entire subgenual functional connectivity map previously generated using random effects analysis across 98 subjects (SG-based seed map). The third seed was the entire functional connectivity map of effective versus ineffective DLPF stimulation sites, previously generated using a paired t-test across 98 normal subjects (Efficacy-based seed map). The later two “seed maps” cover the entire brain with the exception of the left DLPFC which was omitted to avoid biasing of results in this region. Time courses were extracted from each seed region or seed map by computing a weighted average of all voxels within the seed at each time point. Voxel weights for the seed maps were continuous variables and could be positive or negative. The Pearson's correlation coefficient was computed between this extracted time course and that of all other voxels. Fisher's r-to-z transformation was used to convert correlation maps into z maps.
To determine the similarity between two functional connectivity maps, a spatial correlation coefficient was computed across voxels (Fox et al., 2006). To normalize the distributions for statistical comparison, spatial correlation values were converted using Fisher's r-to-z transform. These values are listed in the text as z(r). After averaging and statistical comparison, these values were transformed back to r values using the inverse Fisher transform and the percentage variance accounted for in each comparison was computed as r2 (Supplementary Table 1). Although intuitive, the spatial correlation coefficient may miss some differences when comparing two maps (i.e. scalar offsets or differences in magnitude) so we also compared maps using a spatial eta coefficient (Cohen et al., 2008). Results were nearly identical using these two metrics of similarity, so only results using the spatial correlation coefficient are presented here.
To create a set of potential TMS target sites, we first identified all voxels on the surface of the standard MNI152 brain template based on grey scale intensity (21415 nodes). These voxels were then down-sampled to 4mm resolution (4045 nodes) and masked to include only those voxels in the left DLPFC (163 nodes). The center coordinates for each of these surface nodes were identified resulting in a set of 163 potential TMS target coordinates in the left DLPFC separated from each other by at least 3 mm in all directions (Supplementary Figure 1A). Using in house software, we generated spheres of varying radii centered on these node coordinates to reflect varying focality of different potential cortical stimulation techniques (Supplementary Figure 1B, C). Spheres ranged in size from 1 to 30 mm in radius for analysis purposes, but this is not meant to suggest that commercially available TMS coils are capable of focally stimulating a 1 mm area. Each sphere was masked to exclude voxels outside the brain. For most analyses these spheres were of uniform intensity to simplify the analysis, interpretation, and comparison across sphere sizes. However, for one analysis surface spheres were constructed to have decreasing intensity with increased distance from the sphere center to more accurately model a true TMS coil (Supplementary Figure 2). It is important to note that accurately modeling the volume of TMS-induced changes in neuronal activity is complicated and depends on the specific stimulation coil, intensity, coil orientation, underlying anatomy, and assumed neuronal depolarization function (Deng et al., 2012; Wagner et al., 2007). Such advanced modeling is beyond the scope of the present investigation. To generate a rough approximation, we created a series of concentric spheres of radius 2, 4, 7, 9, and 12 mm (12 mm cone) chosen to match the linear decay of metabolic changes seen in animal TMS experiments (Valero-Cabre et al., 2005). When compared to a model of the induced electric field of a standard figure 8 TMS coil (Pascual-Leone et al., 2002; Thielscher and Kammer, 2004), the distance of 12 mm corresponds to a field intensity of around 75% of maximum (Supplementary Figure 2). This seems appropriate given that this intensity is near the limit at which sub-threshold conditioning pulses produce a physiological effect in human paired-pulse experiments (Kujirai et al., 1993).
The optimal stimulation node for a given functional connectivity map and sphere size was identified by computing the average voxel value (Fisher z score) within each surface sphere then selecting the node with the most negative value. For our 12 mm cone model the varying ROI intensity was normalized to a mean of 1 and a weighted average was used to select the optimal ROI. Using this approach, optimal stimulation nodes were identified for each subject, day, sphere size, and functional connectivity map. For comparison, optimal stimulation nodes were also identified for each sphere size and seed region using the three group-level functional connectivity maps from the independent 98-subject cohort (dataset 1). To compare ROIs selected on the basis of individualized maps to those selected on the basis of the group map, the average voxel value (Fisher z score) within each ROI was computed using the single-subject functional connectivity data from the alternate day (i.e. the day not used to select the stimulation node). The superior targeting technique for a given subject is the one that identified the stimulation node with the more negative value in the independent dataset. The difference between individualized and group based targeting as well as the effect of sphere size was determined using a two-way ANOVA and followed by paired t-tests for each sphere size and seed map.
It is important to note that this reproducibility analysis is not designed to determine if the same stimulation node will be identified in the same subject on two different days, but rather to determine if the stimulation node identified on one day still overlies an anticorrelated area on the second day. This analysis approach is less susceptible to random variation that might occur in a subject with multiple foci of anticorrelation.
All data processing, calculations, and thresholding were performed in volume space. For display purposes data were mapped to the cortical surface using CARET and the PALS atlas (Van Essen, 2005) using the average fiducial mapping option.
Functional connectivity was computed with three seed regions / seed maps to identify candidate TMS targets in the left DLPFC (Figure 2). Regardless of whether one used the small subgenual seed region (Fig 2A), the full subgenual-based seed map (Fig 2B), or the efficacy-based seed map (Fig 2C), a clear anticorrelated area was identified at the group level (black circles). Peak MNI coordinates for these anticorrelations for the three group maps are (−42, 38, 34), (−42, 44, 26), and (−38, 44, 26) respectively. However, when examining results from single subjects, pronounced heterogeneity was apparent. For example, as illustrated in Figure 2, the ideal group-level stimulation target may suffice for Subject 1, but is far from the ideal target for Subject 2.
To obtain an estimate of the magnitude of these individual differences in functional connectivity, the standard deviation across subjects for every voxel in the left DLPFC was computed (Supplementary Figure 3). For reference, this variance was compared to the standard deviation across all left DLPFC voxels within a subject. For the subgenual seed region, the variability across subjects was actually slightly larger than the variability across space within the DLPFC (0.192 vs 0.183 z(r)). Averaged across all three seed regions/maps, the variability across subjects was 0.241 while the variability across space was 0.285 z(r). This suggests that differences in DLPFC connectivity across subjects are of similar magnitude to differences in connectivity across space within the left DLPFC.
To determine if these individual differences were reproducible across time, DLPFC connectivity with the three seed regions / seed maps was computed in 42 subjects scanned on two separate days (Figure 3). The functional connectivity maps for subject one on day one appear similar to the maps from the same subject on day two, but very different from the maps of subject two. This qualitative impression was quantified by computing the similarity (spatial correlation) between DLPFC connectivity maps from the same subject on two separate days versus the similarity between maps from different subjects (Figure 4, left). All three seed regions/maps showed significantly more consistency within a subject than between subjects including the subgenual seed region (z(r)=0.345 vs 0.111, P<10−4), subgenual-based seed map (z(r) = 1.22 vs 0.537, P<10−17) and efficacy-based seed map (z(r) = 1.32 vs 0.567, P<10 −21). These Fisher z values were transformed into r2 values to estimate the percent variance in one map explained by another map and ranged from 1% (z(r) = 0.11) to 75% (z(r) = 1.32) (Supplementary Table 1). DLPFC connectivity remained reproducible within a subject even when comparing across our three different seed regions. Specifically, DLPFC connectivity within a subject with different seeds (and different sessions) was more similar than the same seed across different subjects (z(r) = 0.618 versus 0.396, P<10−6).
The above results attest to reproducibility of individual differences and indicate that subjects are more similar to themselves scanned on a different day than they are to other subjects. However, no one would propose targeting TMS in one subject based on connectivity results from another subject. A more pertinent comparison for determining the best way to target TMS is to compare single subject maps to that of the population (Figure 4, right). Subjects were more similar to themselves scanned on a different day than they were to the population map for the subgenual-based seed map (z(r)=1.22 vs 0.846, P<10−13) and efficacy-based seed map (z(r)=1.32 vs 0.898, P<10−12) but not for the small subgenual seed region (z(r)=0.345 vs 0.344, P>0.9).
Importantly, not all seed regions performed equally well. Both the subgenual-based and efficacy-based seed-maps were significantly more reproducible across days than the smaller subgenual seed region (P<10−16). In fact, connectivity with the subgenual-based seed map was a better predictor of connectivity with the small subgenual seed region on a separate day than connectivity with the subgenual seed region itself (z(r)=0.426 vs 0.345, P < 0.05). This result held true even when the subgenual and the majority of the ventral medial prefrontal cortex was excluded from the subgenual-based seed map (Supplementary Figure 4).
It is important to determine if these individual differences in DLPFC connectivity can actually be translated into individualized TMS targets constrained to the cortical surface. Further, it is important to determine if a TMS target selected on the basis of individualized functional connectivity has the potential to result in a better clinical outcome than a target selected on the basis of functional connectivity averaged across a large group. An example of our analysis approach for addressing these questions is shown for a single subject (Figure 5). An optimal TMS target (6 mm sphere) for this subject was selected both on the basis of group-averaged connectivity (red box) and individualized connectivity from day one (blue box). The potential clinical utility of these targets was then tested using the independent functional connectivity results from that subject for day 2 (black box). Specifically, we determined which target corresponded to the more negatively correlated voxels in the independent dataset. For this subject and sphere size, individualized targeting selected a better (more negatively correlated) stimulation site than population-based targeting, which could lead to a better clinical response.
Similar analyses were conducted for all subjects using our three possible seed regions / maps with stimulation fields ranging in size from 1 mm radius to 30 mm radius (Figure 6). Two-way ANOVAs were conducted for each seed region using targeting method and sphere size as factors. For each seed region / seed map there was a significant difference between individualized and group-based targeting (F = 16.15 to 37.99, df = 1, P < 10−4), a significant difference between sphere sizes (F = 5.88 to 209.56, df = 5, P<10−4), and a significant interaction between the two (F = 4.51 to 5.04, df = 5, P < 0.0005). F values reflect the range across the three seed regions while P values reflect the least significant effect. Follow-up paired t-tests showed that individualized targeting identifies a better stimulation site (i.e. more anticorrelated node) than group-based targeting at several sphere sizes for each seed map (Figure 6). The benefit of individualized targeting becomes more significant as the sphere size decreases and is more significant for the two seed maps compared to the smaller and noisier subgenual ROI. In fact, at the largest (30 mm) sphere size there was actually a significant benefit of using group-based over individualized targeting for the small subgenual seed region.
To obtain a rough approximation of the size of the stimulation field and potential benefit of individualized targeting for commercially available figure 8 TMS coils, we compared individualized to group-based targeting using a 12 mm radius cone-shaped field rather than a simple sphere (Figure 6, column labeled cone). A two-way ANOVA using targeting method and seed region as factors revealed a significant difference between individualized and group based targeting (F = 6.00, df = 1, P<0.05), a significant difference between seed regions (F = 238.4, df = 2, P<10−15), but no significant interaction (F = 0.322, df = 2, P = 0.72). Follow-up paired t-tests showed a significant benefit of individualized targeting for the subgenual-based seed map (P < 0.007) and the efficacy-based seed map (P < 10−4) but not the small subgenual seed region (P = 0.12).
Although normal subjects provide a large and useful dataset for methodological development and statistical testing, it is important to know if the methods have any utility in the population of interest, namely patients with medication-refractory depression presenting for brain stimulation. Repeating the above analyses on two such patients, results replicated the findings seen in the normal subjects. Results for the subgenual-based seed map are displayed here (Figure 7) with the results for other seed regions and node coordinates for each patient tabulated in supplementary material (Supplementary Table 2). Both patients were more similar to themselves scanned on a different day than to the other patient for all three seed regions. Patients were also more similar to themselves than the group map for the subgenual-based and efficacy-based seed maps, but not for the small subgenual ROI. Using the 12 mm cone as our TMS field model, selecting a TMS site based on individualized connectivity identified a more anticorrelated node in the independent fMRI data than selecting the site based on group connectivity.
There are several novel results in the present paper important for successful individualized targeting of TMS to the DLPFC based on functional connectivity. First, individual differences in DLPFC connectivity are large and reproducible across sessions. Second, TMS targets can be selected based on these individual differences and have the potential to be clinically superior to targets selected on the basis of a group map. Finally, individualized targeting might be improved through the use of a seed map over a seed region and becomes particularly advantageous when considering more focal stimulation.
The idea that one should target therapeutic TMS in depression based on individual differences in anatomy or function is not new. Further, it is widely recognized that the conventional 5 cm targeting technique is insufficient (Ahdab et al., 2010; Fitzgerald et al., 2009; Herbsman et al., 2009; Herwig et al., 2001; Padberg and George, 2009). Methods to account for individual differences in anatomy have been proposed including targeting based on standardized EEG electrode positions (Herwig et al., 2003b), or specific MRI coordinates (Fitzgerald et al., 2009; Herbsman et al., 2009; Rusjan et al., 2010). A randomized trial targeting anatomical coordinates chosen based on group-level DLPFC hypometabolism failed to reach its primary endpoint, but did show some clinical benefit beyond the 5 cm approach (Fitzgerald et al., 2009). Taking this approach a step further, three trials have aimed to target TMS based on individualized hypometabolic foci indentified with either SPECT (Garcia-Toro et al., 2006) or PET (Herwig et al., 2003a; Paillère Martinot et al., 2010). All three trials failed to show benefit beyond the conventional 5 cm approach.
A question relevant to the present investigation is why targeting a group-average focus of hypometaboism would be successful while targeting individualized foci would fail. Many explanations are possible, but three will be considered here. First, there may be inaccuracies in neuronavigated TMS leading to differences between the site that is targeted and the site that is actually stimulated. Further work delineating and improving this relationship will be important for any effort at individualized targeting. Second, hypometabolic foci may not be the ideal target for TMS, and perhaps some other property of the DLPFC is responsible for antidepressant response (such as connectivity to deep limbic regions). Given the anticorrelation between limbic regions and the DLPFC even in normal subjects, DLPFC hypometabolism could be secondary to limbic hyperactivity rather than causal in the disease process (Fox et al., 2012a). Finally, and most relevant to the present investigation, individual PET maps may simply be too noisy to serve as a basis for individualized targeting of TMS, and one is better off targeting a population-average focus. Critically, none of these prior trials of individualized targeting was preceded by an analysis of the reproducibility of individualized hypometabolic foci or an investigation into techniques that might optimize their identification.
These pioneering trials of individualized targeting were instrumental in outlining the motivation and approach of tailoring TMS to specific patients. However, they also serve as an example of the challenges of translating a technique from application at the population level to clinical utility in individual subjects. The importance of methodological studies, such as the present work, for technique validation and optimization prior to embarking on a clinical trial should not be underestimated. Based on the results presented here, we propose an outline for individualized connectivity-based targeting of TMS that can be tested in a clinical trial (Figure 8).
Intrinsic (resting state) functional connectivity MRI is a powerful imaging technique that utilizes correlations in spontaneous fluctuations in the blood oxygenation level-dependent (BOLD) signal to assess functional relationships between regions (Biswal et al., 1995; Fox and Raichle, 2007; Van Dijk et al., 2010). This technique has several theoretical and practical advantages for clinical translation (Fox and Greicius, 2010). Accumulating evidence suggests that individual differences in intrinsic connectivity are behaviorally relevant (Baldassarre et al., 2012; Hampson et al., 2006; Koyama et al., 2011; Seeley et al., 2007; van den Heuvel et al., 2009; Zhu et al., 2011) and to some extent reproducible across scanning sessions (Braun et al., 2012; Cohen et al., 2008; Mannfolk et al., 2011; Meindl et al., 2010; Shehzad et al., 2009; Van Dijk et al., 2010; Wang et al., 2011; Zuo et al., 2010). In a recent analysis of reproducibility across 5 MRI scans over 6 months, we found that the DLPFC has some of the largest and most reproducible individual differences in functional connectivity across any area of cortex (Mueller et al., 2012). These results complement known histological variability in this region (Rajkowska and Goldman-Rakic, 1995) and highlight the potential for individualized targeting.
Only a couple prior studies have used intrinsic connectivity to identify individualized TMS targets (Eldaief et al., 2011; Hoffman et al., 2007) (for review see (Fox et al., 2012b)). No study has focused on identification of individualized targets in the DLPFC, designed a systematized approach for TMS target selection, examined the reproducibility of these targets, or investigated factors to improve individualized targeting. These are all novel contributions of the present investigation and are critical steps towards making individualized connectivity-based guidance of TMS clinically relevant.
A potentially important methodological development presented here is the use of weighted seed maps, rather than small seed regions, to improve signal to noise in single-subject connectivity analyses. If one is attempting to identify a node in the DLPFC anticorrelated with the subgenual in a specific subject, the optimal approach is not to simply perform functional connectivity with the subgenual. Rather, more reproducible results are obtained by first computing functional connectivity with the subgenual on a large independent cohort to generate a seed map. One can then subtract the area of interest from the seed map (in this case the DLPFC) then use the seed map rather than the small subgenual seed region to assess functional connectivity in the single subject. Somewhat surprisingly, the result is a better prediction of subgenual connectivity than would be obtained using the subgenual seed region itself. This approach even works when the subgenual and surrounding medial prefrontal cortex is removed from the seed map, an important finding given the variability with which different scanners and MRI sequences sample this area of high susceptibility (Ojemann et al., 1997). Further work is needed to determine if this approach will be useful in other instances where single-subject functional connectivity with a small or noisy seed region is desired.
Another important finding from the current study is the effect of modeled stimulation field size on the benefits of individualized versus population-based targeting. The smaller the modeled volume of TMS-induced activation, the more important individualized targeting becomes. This factor could become critical for more invasive and potentially focal DLPFC stimulation techniques such as epidural or subdural stimulation (Kopell et al., 2011). Conversely, individualized targeting becomes less important with larger stimulation field sizes. Given this finding, one might reasonably ask what the importance of individualized targeting might be in the current clinical setting with commercially available TMS coils. This is a difficult question to answer as it remains unclear how focal these coils truly are with respect to the volume of activated tissue. Our model of TMS-induced activation using a 12 mm radius cone suggests that individualized targeting may be relevant for current figure 8 coils, however this model was based on several assumptions (see methods) and should be interpreted with caution. Even if the focality of current TMS coils were clearly established, it remains unclear if intense stimulation specifically focused on the most anticorrelated area of cortex or broader stimulation to a wide area of cortex with more weakly anticorrelated voxels would have a better clinical effect. In the end, there are multiple unknown variables regarding the physiological and clinical effects of TMS and true validation of the individualized connectivity-based targeting techniques presented here will require a clinical trial.
First, the current paper is a technique development paper, not a clinical trial. These results support the potential utility of individualized targeting of TMS based on connectivity and may help enable a clinical trial, but the clinical validity of this approach remains to be proven. Second, we used highly simplified models of the tissue volume activated by TMS. More advanced models taking into account coil geometry, coil orientation, stimulation intensity, individual differences in anatomy, and probability of neuronal activation are likely to be important. Third, this paper was focused specifically on left DLPFC targets for depression. The reproducibility of individualized targets in other cortical regions or with alternate deep nuclei awaits similar validation. Finally, proof of concept was demonstrated in only two patients with medication refractory depression. Combined with our prior results (Fox et al., 2012a), this strongly suggests that the methods presented here will be applicable in this patient population, however this awaits statistical validation in a larger patient cohort.
There is significant individual variability in the connectivity of the left DLPFC. This variability is stable across scanning sessions and can be used to generate individualized and reproducible TMS targets. Seed maps demonstrate more stability than a small subgenual seed region and may be an effective technique for improving signal to noise in single subject functional connectivity analyses. Finally, the more focal the stimulation field, the greater the benefit likely to come from individualized targeting. The methods presented here set the stage for a clinical trial of individualized connectivity-based targeting of left DLPFC TMS for depression. Further, these techniques might be applicable to identification of individualized targets for focal brain stimulation across a variety of disorders and stimulation techniques.
Creation of potential stimulation nodes across the left DLPFC and modeling of stimulation volumes with simple spheres. Sagittal brain sections show some of the 163 nodes across the cortical surface of the left DLPFC used to generate potential TMS targets (A). Spheres ranging in size from radius of 1 mm to 30 mm are shown centered on one example node in volume space (B) and on the cortical surface (C).
Creation of the 12 mm cone model of TMS activation. A series of concentric spheres of radius 2, 4, 7, 9, and 12 mm were created and overlapped to generate a cone-shaped model of TMS-induced neuronal activation. This ROI is shown in volume space (A) and on the cortical surface (B) although the exact location of the ROI in the DLPFC was allowed to vary based on the functional connectivity of individual subjects (see methods). The size and linear rate of decay of this model (C, red line) were chosen to match experimental findings in animal models of TMS-induced metabolic changes (C, blue line) and correspond well with electric field models of figure 8 TMS coils thought to induce physiological effects up to ˜75% of maximum intensity. Because electric field models for figure 8 coils differ depending on whether one moves away from the coil in depth (C, light green line) or laterally along the surface (C, purple line), these two decays were averaged (C, green line) to assist in modeling of a spherical stimulation volume.
Variability in connectivity across subjects is similar in magnitude to variability in connectivity across space in the dorsal lateral prefrontal cortex (DLPFC). Shown is the standard deviation across subjects for all voxels in the left DLPFC (individual variability, white bars) compared to the standard deviation across voxels for all subjects (spatial variability, black bars) for the three seed regions / seed maps considered in the present article.
Benefit of using a distributed seed map for predicting subgenual connectivity. Connectivity with the subgenual based seed map (SGmap, B) better predicted connectivity with the small subgenual ROI (SG, A) on a different day than connectivity with the small subgenual ROI itself. This remained true even when most of the ventral medial prefrontal cortex was excluded from the subgenualbased seed map (SGmap_subMPF, C).
We thank the Brain Genomics Superstruct Project for contributing data and David Alsop for assistance with MRI acquisition and sequence optimization for the depression patients. MDF was supported by NIH grant R25NS065743 and HL by NIH grant K25NS069805. APL serves on the scientific advisory boards for Nexstim, Neuronix, Starlab Neuroscience, Neuroelectrics, Neosync, and Novavision, and is listed as inventor in issued patents and patent applications on the real-time integration of transcranial magnetic stimulation (TMS) with electroencephalography (EEG) and magnetic resonance imaging (MRI). Work on this study was also supported by grants from the National Institutes of Health and National Center for Research Resources: Harvard Clinical and Translational Science Center (UL1 RR025758).
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.