Functional Connectivity Mapping for rTMS Target Selection in Depression

Abstract

Objective: Repetitive transcranial magnetic stimulation (rTMS) protocols increasingly use subgenual anterior cingulate cortex (sgACC) functional connectivity to individualize treatment targets. However, the efficacy of this approach is unclear, with conflicting findings and varying effect sizes across studies. Here, the authors investigated the effect of the stimulation site's functional connectivity with the sgACC (sgACC-StimFC) on treatment outcome to rTMS in 295 patients with major depression.

Methods: The reliability and accuracy of estimating sgACC functional connectivity were validated with data from individuals who underwent extensive functional MRI testing. Electric field modeling was used to analyze associations between sgACC-StimFC and clinical improvement using standardized assessments and to evaluate sources of heterogeneity.

Results: An imputation-based method provided reliable and accurate sgACC functional connectivity estimates. Treatment responses weakly but robustly correlated with sgACC-StimFC (r=-0.16), but only when the stimulated cortex was identified using electric field modeling. Surprisingly, this association was driven by patients with strong global signal fluctuations stemming from a specific periodic respiratory pattern (r=-0.49).

Conclusions: Functional connectivity between the sgACC and the stimulated cortex was correlated with individual differences in treatment outcomes, but the association was weaker than those observed in previous studies and was accentuated in a subgroup of patients with distinct, respiration-related signal patterns in their scans. These findings indicate that in a large representative sample of patients with major depressive disorder, individual differences in sgACC-StimFC explained only ∼3% of the variance in outcomes, which may limit the utility of existing sgACC-based targeting protocols. However, these data also provide strong evidence for a true-albeit small-effect and highlight opportunities for incorporating additional functional connectivity measures to generate models of rTMS response with enhanced predictive power.

Keywords: Functional connectivity; Major depressive disorder; Neuroimaging; Neurostimulation; Subgenual anterior cingulate cortex; Transcranial magnetic stimulation.

FIGURE 1.. Reliability and accuracy of determining subgenual anterior cingulate cortex (sgACC) functional connectivity (FC) with the weight-map method^a

^a We used multiecho fMRI data (ME) to investigate how accurately the weight-map method approximates true sgACC functional connectivity (FC). Panel A shows cortical surface maps depicting sgACC FC for five densely sampled subjects (ME01–ME05), each with up to 15 hours of concatenated multiband multiecho fMRI data. Because the simple sgACC seed yielded reliable FC maps in the concatenated multiecho fMRI data (22), we were able to test how accurately the FC maps derived from the weight-map approach (right column) approximated the “ground truth” sgACC seed-derived FC maps (left column). Qualitatively, there was a high correspondence between the FC maps generated by the two methods. The median spatial correlation (r) between the resulting maps for the two methods was 0.95. As shown in panel B, we next tested the performance of the weight-map method with single-echo fMRI data from the publicly available Midnight Scan Club (MSC) data set (21) that consists of data from 10 densely sampled individuals, each with 10 30-minute single-echo fMRI scans. Cortical surface maps depict FC for all MSC subjects from concatenated time series (5 hours) derived from either a simple sgACC seed (left column) or the weight-map method (right column). Maps derived from the sgACC seed were very noisy, whereas the weight-map method yielded FC maps with a consistent default mode network-like network configuration resembling the multiecho results. Note that the weight-map method produced higher absolute FC values with both the MSC and ME data sets. Panel C shows the test-retest reliability in terms of the mean spatial correlation (r) across each subject’s 10 fMRI sessions for a simple spherical sgACC seed (red) and the weight-map method (blue). Note the consistent improvement in reliability with the weight-map method across subjects, with a mean improvement from r=0.38 to r=0.81 for the total sample. Panels B and C show results for only five of 10 MSC subjects; results of the entire sample are shown in Figure S4 in the online supplement.

FIGURE 2.. Subgenual anterior cingulate cortex (sgACC) functional connectivity (FC) with the stimulation site predicts individual differences in treatment outcomes^a

^a Panels A–B show in a representative subject how the individual rTMS stimulation coordinates on an inflated cortical surface (panel A) relate to the relative distribution of the induced electric field (E-field) (panel B, image at left), with the cortical area receiving the 99th percentile strongest E-field outlined in black. The image at right in panel B illustrates how the stimulated area of the dorsolateral prefrontal cortex (DLPFC) (with the 99th percentile strongest E-field outlined in black) maps onto the same subject’s sgACC FC map (derived with the weight-map approach [2,31]). As shown in panel C, sgACC FC with the stimulation site as estimated by E-field modeling was negatively correlated with treatment outcomes (% improvement in total score on the 16-item Quick Inventory of Depressive Symptomatology [QIDS-SR]) (r=−0.16, p=0.006). Panels D and E show the corresponding results for the weighted-cone approach, in which a 12-mm distance-weighted hemisphere around the target coordinate is used to average FC features (2). Panel D (image at left) illustrates the borders of a 12-mm weighted cone on the inflated cortex, where the distance to the stimulation target site is color coded, and shows results markedly different from those of the projected E-field estimation. The apparent asymmetry of the radius is due to the projection from a folded cortex to a flattened cortex. The image at right in panel D illustrates how the 12-mm weighted cone maps to the sgACC-StimFC within the DLPFC. As shown in panel E, sgACC FC with the stimulation site as estimated by the weighted-cone approach was not associated with treatment outcomes (r=−0.03, p=0.56); x-axis values denote weighted averages.

FIGURE 3.. Sources of effect size variance^a

^a Panel A shows the effect of sample size on effect size variance in 10,000 bootstrapped analyses per sample size. In samples of 25 subjects, the median sample size in previous studies, effect sizes (r) of −0.5 are often found by chance. As shown in panel B, functional connectivity of the stimulated site with the sgACC (sgACC-StimFC) is highly predictive of clinical improvement in the subset of subjects with the lowest temporal signal-to-noise ratio (tSNR) (orange arrow). For this analysis, subjects were first ordered by z-scored tSNR values. Each marker represents the correlation strength between sgACC-StimFC and clinical improvement in a subsample of 50 subjects. Samples were then drawn in ascending order, based on tSNR values in an overlapping sliding window manner. Gray shading represents 95% of the null distribution of r values in 10,000 random draws of 50 subjects. No dependence of the effect on movement (mean framewise displacement) was found (see Results section in the online supplement).

FIGURE 4.. Association between functional connectivity of the stimulated site with the subgenual anterior cingulate cortex (sgACC-StimFC) and clinical improvement is carried by a subpopulation of subjects with high breathing-specific signal variance^a

^a As shown in panel A, low temporal signal-to-noise ratio (tSNR) in our sample was explained by differences in signal variance (image at right) as opposed to differences in mean intensity of the fMRI signal (image at left). Panel B shows that a major source of variance in the global fMRI signal stems from breathing (20). Three blinded raters scored all 590 fMRI scans (two per subject) for the presence of two common irregular breathing patterns: burst breathing and deep breaths. Representative examples of carpet plots (rows are voxels, and columns are time points; intensity represents signal strength) are shown for normal breathing, burst breathing, and deep breaths of individuals included in the Human Connectome Project (HCP) data set (used and reproduced from [20]) and the THREE-D data set. Note the strong correspondence of characteristic patterns that are clearly identifiable without a breathing-belt trace. Panel C shows the high interrater reliability that was obtained between each pair of raters for each of the breathing patterns (Cohen’s kappa and r; table at left). A sex bias was found for the occurrence of burst breathing but not deep breaths (figure at right), which exactly replicated a previous study (20). Panel D shows that among the 47 subjects showing signal fluctuations indicative of burst breathing, sgACC-StimFC was highly predictive of clinical improvement (% improvement on the 16-item Quick Inventory of Depressive Symptomatology [QIDS-SR]) (figure at left). The probability of finding an effect of this strength in a random subsample of 47 subjects was p=0.002 in 10,000 bootstrapped subsamples. Gray vertical bars present nominally significant effect sizes, and the green vertical bar marks the actual effect size (figure at right). Panel E shows that when these 47 subjects with signal fluctuations indicative of burst breathing were removed from the total sample, no significant association between sgACC-StimFC and clinical improvement remained (figure at left). The strong association was specific to samples that included those showing burst breathing and not present in a control sample of 47 subjects with equally low tSNR values (105.56 vs. 117.48 in those showing burst breathing) that did not show evidence for burst breathing (figure at right). Conversely, when this control sample was removed from the total sample, the association between sgACC-StimFC and clinical improvement was unchanged (r=−0.14, p=0.03; data not shown). As shown in panel F, the dependency of the association between sgACC-StimFC and clinical improvement on tSNR was removed in the absence of the 47 subjects with evidence for burst breathing, indicating that this phenomenon was driven by the presence of individuals showing burst breathing in the sample rather than being a consequence of nonspecific sources of high signal variance. One subject with a QIDS-SR improvement of −150% was omitted for illustrative purposes; the subject was included in all test statistics.