Precision Network Modeling of Transcranial Magnetic Stimulation Across Individuals Suggests Therapeutic Targets and Potential for Improvement

Abstract

Higher-order cognitive and affective functions are supported by large-scale networks in the brain. Dysfunction in different networks is proposed to associate with distinct symptoms in neuropsychiatric disorders. However, the specific networks targeted by current clinical transcranial magnetic stimulation (TMS) approaches are unclear. While standard-of-care TMS relies on scalp-based landmarks, recent FDA-approved TMS protocols use individualized functional connectivity with the subgenual anterior cingulate cortex (sgACC) to optimize TMS targeting. Leveraging previous work on precision network estimation and modeling of the TMS electric field (E-field), we asked whether various clinical TMS approaches target different functional networks between individuals. Results revealed that modeled homotopic scalp positions (left F3 and right F4) target different networks within and across individuals, and right F4 generally favors a right-lateralized control network. TMS coil positions over the dorsolateral prefrontal cortex (dlPFC) zone anticorrelated with the sgACC most frequently target a network coupled to the ventral striatum (reward circuitry) but largely miss that network in some individuals. We further illustrate how modeling can be used to retrospectively assess the estimated targets achieved in prior TMS sessions and also used to prospectively provide coil positions that can target distinct closely localized dlPFC network regions with spatial selectivity and maximal E-field intensity. In a final study, precision targeting was found to be feasible in participants with Major Depressive Disorder using data derived from a single low-burden MRI session suggesting the methods are applicable to translational efforts where limiting patient burden and ensuring robustness are critical.

Keywords: depression; functional MRI; networks; precision modeling.

FIGURE 1

Left F3 versus right F4 TMS coil positions target distinct networks in a representative participant. Simulation results from one participant are shown with positioning of the TMS coil at left F3 and right F4 sites (homotopic scalp locations based on the Okamoto 10–20 EEG coordinate system). (A) Network estimates on the left and right hemispheres of the cerebral cortex are presented. The network composition within the association zones varies between hemispheres including within dlPFC. The legend below labels the networks: SMOT‐A, Somatomotor‐A; SMOT‐B, Somatomotor‐B; PM‐PPr, Premotor‐Posterior Parietal Rostral; CG‐OP, Cingulo‐Opercular; SAL, Salience; dATN‐A, Dorsal Attention‐A; dATN‐B, Dorsal Attention‐B; FPN‐A, Frontoparietal Network‐A; FPN‐B, Frontoparietal Network‐B; DN‐A, Default Network‐A; DN‐B, Default Network‐B; LANG, Language; VIS‐C, Visual Central; VIS‐P, Visual Peripheral; AUD, Auditory. (B) Maps of the E‐field effects are displayed for the left F3 and right F4 sites. Blue colors represent lower and red colors higher E‐field (V/m) values. (C) Overlap between the maps of the E‐field model and network estimates quantified at various E‐field thresholds are plotted. At left F3, the top 1%–0.1% (99.0%–99.9%) of the E‐field includes multiple networks without a clear predominant network. At right F4, there is relatively more FPN‐B in the E‐field, at each threshold. (D) Distribution of E‐field intensity values across the estimated networks with the modeled stimulation dose set at dI/dt = 48 A/μs. Compared to other networks, models suggest there is higher stimulation intensity supplied to FPN‐A and CG‐OP on the left, and FPN‐A and FPN‐B on the right. The left hemisphere receives higher overall intensity than the right hemisphere.

FIGURE 2

Selectivity and intensity for left F3 versus right F4 TMS coil positions quantified in 15 participants. Plots of selectivity and intensity illustrate the difference between the modeled effects of scalp landmark‐based left F3 and right F4 stimulation on the FPN‐B versus SAL networks. Each participant is represented by one pair of connected symbols corresponding to the TMS coil placements: Left F3, gray circle; Right F4, black triangle. The left panel displays the selectivity corresponding to the relative % of SAL and FPN‐B in the top 0.5% of the E‐field. The right panel displays the intensity of the highest 25 vertices for FPN‐B and SAL (each point is the mean of the highest 25 vertices). While the pairs of connected dots show an overall horizontal pattern reflecting that SAL is similarly targeted across hemispheres, the rightward shift for the right F4 estimates indicates that FPN‐B, a right‐lateralized candidate control network, is targeted more by positioning the TMS coil on the right hemisphere. Note also that the degree of SAL selectivity and intensity varies between participants considerably for both sites.

FIGURE 3

The sgACC anticorrelation strategy targets multiple networks with variability between individuals. Plots illustrate two facets of the sgACC anticorrelation targeting strategy. (A) The region targeted by the sgACC anticorrelation contains subregions linked to multiple distinct networks. The plot displays the % of vertices in each network within the anticorrelated target region in 15 participants, with each bar representing the mean for a single network and the circles representing individual participants. The sgACC anticorrelated target region variably includes SAL, CG‐OP, FPN‐A, and dATN‐A network regions. Note specifically the marked variability between individuals in the degree to which the SAL network is included. (B) Selectivity and intensity at the individualized sgACC anticorrelated target are quantified and plotted. Circles represent the 15 individual participants. The left panel shows selectivity of the anticorrelated target, quantified as relative % of FPN‐A versus SAL within the thresholded E‐field (top 0.5% of values). The right panel shows the maximal intensity (mean of the 25 highest values) at the anticorrelated target for FPN‐A versus SAL. The five highlighted participants (1–5) demonstrate that the degree of selectivity does not always predict the values for maximal intensity.

FIGURE 4

Precision network mapping and E‐field modeling can be prospectively applied to target networks in individuals. Adapted views from the report generated by the precision TMS pipeline are displayed for one typical participant. (A) Network estimates are displayed on the native‐space surface with colors denoting distinct networks. The network colors are the same as in Figure 1. (B) The dlPFC region used to constrain the search space for target selection is displayed in green. Note that the region overlaps with many distinct networks in panel A. (C) The largest continuous cluster on a gyral crown within dlPFC is selected for each of the targets (highlighted in black) on top of the target network estimates shown in color. Note that an isolated network (e.g., FPN‐A, left) or network pairs (e.g., DN‐A & DN‐B, right) can be set as targets, which illustrates the flexibility of this approach. (D) The E‐field map corresponding to the best coil placement for each target is shown. Blue colors represent lower and red colors higher E‐field (V/m) values. (E) The overlap between the E‐field map and estimated networks is quantified at multiple thresholds (99.0%–99.9%). This participant demonstrates high spatial selectivity for each of the targets that increases as the thresholds increase. (F) Plots show the distribution of E‐field values within networks, with the modeled stimulation dose set at dI/dt = 48 A/μs. Values on the right of the plot indicate high E‐field intensity, which is theoretically supplied to FPN‐A in this participant, but not SAL & CG‐OP or DN‐A & DN‐B. The E‐field intensity plot considers all vertices in the cortex including those at depth and those outside the search space. It thus represents the estimated stimulation effect for the TMS site and is independent of search space and target selection assumptions.

FIGURE 5

Hypothetical spatial selectivity and E‐field intensity achievable in 15 participants. Plots display the modeled spatial selectivity and E‐field intensity that is achieved by optimizing coil positions within individuals. Each pair of connected symbols represents one participant, and colors correspond to the optimal TMS coil placement for distinct networks: FPN‐A, orange; SAL & CG‐OP, purple; DN‐A & DN‐B, maroon. Left panels show the selectivity (relative % in the E‐field at the top 0.5% of values); right panels show the maximal intensity of the highest 25 vertices (each point is the mean of the highest 25 vertices). The top row compares FPN‐A versus SAL & CG‐OP; the middle row compares FPN‐A versus DN‐A & DN‐B, and the bottom row compares SAL & CG‐OP versus DN‐A & DN‐B. Overall, there is clear separation in selectivity between the optimized coil positions, such that the coil placed at the target network(s) is more selective for the target than non‐target network(s). This is also observed for maximal intensity, though the separation is narrower again reminding that the effects of selectivity and intensity can be distinct, and both should be taken into account.

FIGURE 6

Higher‐order network estimates from a single MRI session are replicated in independent multi‐session data. Networks targeted by TMS, as shown here, were modeled using a single session of data (left) versus prospective data independently acquired over 10 sessions within the same individuals (right). The legend below labels the networks: CG‐OP, Cingulo‐Opercular; SAL, Salience; FPN‐A, Frontoparietal Network‐A; DN‐A, Default Network‐A; DN‐B, Default Network‐B. The black outline region in each panel is the target SAL/CG‐OP region used for TMS targeting as estimated from each individual's single‐session baseline acquisition (left). The exact region outline is projected and shown on the independent network estimates from the prospective data (right). Note that the network estimates are quite stable between the independent data samples and further that the SAL/CG‐OP target region estimated solely from the single‐session data predicts extremely well the location in the independent data. Similar results are shown for all participants in Figures S19 and S21.

FIGURE 7

Precision network mapping and E‐field modeling prospectively applied to target networks using single‐session data in a participant with MDD. Mirroring the structure of Figure 4, using a single session of resting‐state data (~1 h), the precision TMS pipeline was used to prospectively target distinct networks in a participant with MDD. (A) Network estimates shown on the native‐space surface. (B) The dlPFC search space used for target selection. (C) The network target regions (highlighted in black), overlayed on network estimates in color. (D) The E‐field map corresponding to the best coil placement for each target. Blue colors represent lower and red colors higher E‐field (V/m) values. (E) The overlap between the E‐field map and estimated networks quantified at multiple E‐field thresholds (99.0%–99.9%). This participant demonstrates preferential spatial selectivity for each of the targets that increases as the thresholds increase. (F) Distribution of E‐field values within networks, with the modeled stimulation dose set at 120% of this participant's resting motor threshold (dI/dt = 49 A/μs). Values on the right of the plot indicate high E‐field intensity, which is theoretically supplied to FPN‐A and SAL & CG‐OP in this participant, but not DN‐A & DN‐B.

FIGURE 8

Achieved spatial selectivity and E‐field intensity in eight participants. Mirroring the structure of Figure 5, these plots display the achieved spatial selectivity and E‐field intensity during real‐world TMS sessions. Each symbol represents one participant, each pair of connected symbols represents the same TMS session number in the sequence of 3 TMS sessions administered per target, and colors correspond to the optimal TMS coil placement for distinct networks: FPN‐A, orange; SAL & CG‐OP, purple; DN‐A & DN‐B, maroon. Left panels show the selectivity (relative % in the E‐field at the top 0.5% of values); right panels show the intensity of the highest 25 vertices (each point is the mean of the highest 25 vertices). Overall, there is clear separation in selectivity and maximal intensity between the optimized coil positions, such that the coil placed at the target network(s) is more selective and achieves higher intensity for the target than non‐target network(s).