Introduction

Transcranial magnetic stimulation (TMS) to the dorsolateral prefrontal cortex (DLPFC) is an effective treatment for major depressive disorder (MDD), but response rates vary significantly, ranging from 30% to 80% across randomized [1, 2] and naturalistic studies [3, 4]. This variability has driven development of novel TMS approaches, exploring optimized targeting methods [5,6,7], alternative stimulation protocols [8, 9], or innovative coil designs [10]. One such development is the H1 coil, which was cleared by the US Food and Drug Administration since 2013 after an international multi-center clinical trial [11], and is now one of the most common TMS devices in clinical practice [12].

With a figure-of-8 coil, some evidence suggests that the efficacy of TMS might depend on choosing the ideal brain circuit to target [7, 13,14,15,16,17]. For example, recent combined data from brain lesions that cause depressive symptoms, from TMS and deep brain stimulation (DBS) sites that modify depression severity [16], and later from electroconvulsive therapy [17], revealed a common causal circuit for depression, suggesting that modulating this network might optimize symptom improvement. However, the H1 coil stimulates the DLPFC more broadly and deeply than figure-of-8-coils [18]. It could be argued that an electric field (E-field) as wide as 18 cm3 of brain volume (compared to 3 cm3 for the figure-of-8 coil) [19] may be too broad to target specific circuits. Conversely, it is also plausible that the larger generated E-field may better engage brain networks spanning large cortical and subcortical regions. For these reasons, it remains unclear if circuit-based targeting is feasible with a deep H1 coil.

In this study, we investigated whether effectiveness of TMS with an H1 coil might be linked to incidental targeting of a specific brain circuit. We retrospectively analyzed a sample of 97 MDD patients treated with the H1 coil, who had previously completed a structural MRI scan for unrelated reasons. Due to inter-individual variability in brain and scalp anatomy, we anticipated that different brain regions would be incidentally targeted in different patients. Using E-field modeling and normative resting-state functional connectivity, we aimed to determine: 1) whether clinical outcomes depend on stimulation of specific brain regions, 2) whether stimulation sites that preferentially improve depression are connected to a distinct brain circuit, and 3) whether this circuit is similar to the previously identified causal depression circuit [16]. In post-hoc analyses, we identified an “ideal” stimulation target maximally overlapping with the depression circuit.

Materials and Methods

Sample and TMS procedure

This study was approved by the Mass General Brigham institutional review board. Data were retrospectively collected from 1423 patients receiving TMS for treatment-­resistant depression at McLean Hospital (Belmont, Massachusetts) between 2014-2022. TMS was administered using the H1 Coil (BrainsWay, Jerusalem, Israel) at 18 Hz in 55 trains of 2 s duration, inter-train interval of 20 s, 1980 pulses per session at 120% of resting motor threshold. The stimulation target was localized by moving the coil along the scalp surface 6 cm anterior to the site of the motor cortex (M1) evoking a maximal finger twitch. When this location was either below eyebrow line or created pain, the coil was moved posteriorly by 1 cm for tolerability. Patients completed the self-report Quick Inventory of Depressive Symptomatology (QIDS-SR) [20] at baseline, every 10 treatments, and at the final session. Inclusion criteria required minimum three QIDS assessments including baseline (i.e., 19–20 sessions depending on the day of completion), as ~20 sessions is usually considered the minimum for a therapeutic response [4, 21, 22]. A full 36 sessions were not required for inclusion to mitigate selective dropout bias. Patient charts were manually reviewed to determine if an anatomical T1-weighted MRI scan was incidentally present. Patients with an available T1 of sufficient quality (i.e., no excessive grain, no cropping, no single slices, no loading errors) were included in the study. In order to maximize our sample size, no restrictions on MRI acquisition date relative to treatment onset were imposed. The quality assessment process was blinded to clinical outcomes.

E-field modeling and functional connectivity analyses

A representation of the modeling pipeline is shown in Fig. 1. T1-weighted images were preprocessed using SynthSR, a deep learning tool interpolating 1 mm isotropic MPRAGE images from clinical scans [23, 24]. We performed E-field modeling using SimNIBS v4.0.1 (https://simnibs.github.io), employing a validated model of the H1 coil provided by BrainsWay [25] (Figure S1). Each patient’s preprocessed T1 was first segmented to generate an head model [26]. In prior work, TMS targets 6 cm anterior to M1 on the scalp were found to be approximately 50 mm anterior to M1 on the cortex [27]. Because no neuronavigation or individual TMS site marking was performed, we cross-registered group-based M1 coordinates (MNI:[-42,-16,65]) [27] onto the patient’s own T1, and centered the coil 50 mm anterior to M1. The simulated E-field magnitude volumes were then normalized to MNI space for subsequent analyses. To account for shifts in coil position and validate our positioning approach, we repeated the analyses centering the coil at 40, 45, 55 and 60 mm anterior to M1. We hypothesized that locations other than the actual target would yield weaker results. See Supplementary Methods for detailed description of coil positioning.

Fig. 1: Schematic representation of modelling pipeline.
figure 1

A Each patient’s T1-weighted brain scan was preprocessed to interpolate a 1 mm isotropic standard MPRAGE research scan from a clinical scan. B The preprocessed T1 was segmented to create a 3D head model. Coil center coordinates were determined by transforming MNI coordinates of M1 into each patient’s native space, and assuming that stimulation was applied 50 mm anterior to M1. C A computational model of the H1 coil was used for electric (E-)field simulation at the identified target location. D The magnitude of the simulated E-field volume of each patient was thresholded to extract the strongest E-field. This was used as seed for seed-based functional connectivity analyses: across 1000 subjects from a normative connectome, the Fisher-z transformed Pearson correlation was computed between the seed and every other voxel in the brain. For each patient, an average map was created across the 1000 subjects. Spearman’s correlation was used to correlate clinical outcome with thresholded and unthresholded E-field magnitude volumes, and average normative E-field connectivity maps (shaded green areas across C and D). magnE: magnitude of electric field; MNI: Montreal Neurological Institute.

Full size image

For each patient, we thresholded the resulting E-field magnitude volume to include only the top 1% of voxels. This threshold was selected as it corresponds to an average brain volume of 18.28 cm3, closely aligning with previous estimates that the H1 coil primarily stimulates up to 18 cm3 of brain volume [28]. We extracted peak and mean E-field magnitude within the thresholded E-field, and estimated the distribution of brain regions primarily stimulated, by adding the individual binarized thresholded E-fields to quantify the spatial overlap across patients. Thresholded E-fields were correlated with clinical outcome, and used as seeds in functional connectivity analyses (see paragraph below). To ensure that the results were not threshold-dependent, we repeated the analyses using an unthresholded E-field model, and the subsequent functional connectivity analyses using seeds at more liberal thresholds of 5% and 10% E-field magnitude.

For functional connectivity analyses, we used a normative connectome database (n = 1000) [6, 14]. Seed-based functional connectivity maps were computed as the Fisher-z-transformed Pearson’s correlation between the seed’s average timeseries and each individual voxel’s timeseries. This yielded a whole brain connectivity map for each patient, which was then correlated with clinical outcome (see Statistical Testing). Throughout the manuscript, “E-field connectivity” refers to the whole-brain connectivity of the E-field thresholded at the top 1%, unless otherwise specified.

Statistical testing

Partial correlation was performed at each brain voxel to correlate the E-field magnitude and E-field connectivity values with TMS-induced change in QIDS score. Age, sex, and baseline QIDS were included as covariates. To account for non-normality in the distribution of QIDS difference scores (Kolmogorov-Smirnov test: p = 3.3 × 10-53), Spearman’s correlation was used for all analyses considering clinical outcome. Statistical significance of the outcome correlation maps was determined using Westfall-Young permutation testing for family wise error (FWE) correction [29], randomly reassigning each patient’s clinical outcome to a different patient’s E-field magnitude/connectivity map. After 10,000 iterations, significance was determined if the outcome was stronger for the real data than for 95% of random permutations.

To quantify the explained clinical variance, we performed leave-one-out cross-validation. The outcome correlation maps were recomputed after excluding one participant, and these leave-one-out maps were spatially correlated to the left-out participant. This was repeated for every participant. The leave-one-out correlations were correlated with the actual clinical outcome of the patients, controlling for the same covariates.

Next, we computed the spatial correlation between our results and the a priori convergent depression circuit derived by Siddiqi et al. (2021) [16] combining data on lesions causing depression, TMS and DBS sites modifying depression. This correlation was tested for significance using a permutation test in which each patient’s outcome was shuffled against a different patient’s neuroimaging. The maps were considered similar if the real spatial correlation was stronger than at least 95% of permuted values.

We quantified the clinical variance explained by the convergent depression circuit by computing the spatial correlation between each participant’s E-field connectivity maps and the depression circuit. This was correlated with QIDS change, controlling for the same covariates. We hypothesized that greater E-field connectivity similarity to the depression circuit would predict greater improvement.

Post-hoc analyses

To estimate the magnitude of the clinical effect of targeting the convergent depression circuit, we categorized patients based on whether their E-field connectivity showed ‘high’ (i.e., above group median) or ‘low’ (i.e., below group median) similarity to the depression circuit. Unpaired t-test was used to compare mean QIDS improvement between groups, hypothesizing that patients with high E-field connectivity similarity to the depression circuit would show greater improvement than those with low similarity.

To test for specificity, we correlated E-field connectivity similarity to the depression circuit with change on the 24 items from the Behavior and Symptom Identification Scale (BASIS-24) [30], which assesses mental health outcomes not specific to depression. Age, sex and baseline BASIS-24 item score were used as covariates. We hypothesized that E-field connectivity similarity to the depression circuit would predict improvement in depression more than any other symptom.

Finally, we identified a potential optimized target as the E-field yielding the highest overlap with the depression circuit, similarly to Lynch et al. (2022) [31]. Overlap was quantified by averaging the E-field’s top 1% of voxels multiplied by the corresponding voxel value in the depression circuit. We performed E-field modeling at different locations within the dlPFC on a standard MNI brain. The simulated locations were arranged on a grid centered 45 mm anterior to M1, with the coil center shifted in 5 mm increments in each direction. This range was established to respect dlPFC boundaries while ensuring an acceptable distance from M1.

Results

Ninety-seven patients (50 males, mean age 43.39yo (±18.86)) completed 19-36 treatment sessions with full symptom scores, and had a T1-weighted MRI scan of sufficient quality (average resolution: 0.92(±0.08) x 0.94(±0.06) x 1.12(±0.81)). Patient demographics, baseline clinical characteristics and main reported indications for undergoing a brain MRI are reported in Table S1. On average, baseline depressive symptoms were severe (mean baseline QIDS: 20.22 (±6.33)), and showed a 34.83% (±28.28%) reduction with treatment (mean QIDS difference: 6.94 (±6.19), Figure S2). We recorded a response and remission rate of 28.87% and 10.31%, respectively.

Correlation between E-field magnitude and outcome

There was no significant association at any voxel between thresholded E-field magnitude and clinical outcome after multiple comparisons correction (pFWE>0.431) or leave-one-out cross-validation (r = −0.010, p = 0.926). The strongest uncorrected correlation was observed in the right medial superior frontal gyrus (r = −0.395, MNI: [12,54,30]), indicating smaller improvement when stimulation at this location was strongest (Fig. 2A, B). No association was found with peak or mean E-field magnitude within thresholded E-field (Figure S3).

Fig. 2: Clinical outcome is not significantly related to stimulation of brain regions when not considering the underlying brain circuit.
figure 2

A Uncorrected voxel-wise correlation map of thresholded E-field magnitude volume (including the top 1% of voxels) and QIDS improvement (calculated as the difference between pre- and post-treatment scores), controlling for age, sex and baseline symptoms. The correlation peak was at MNI coordinate [12,54,30] (red crosshair), but did not survive multiple comparisons correction. B E-field magnitude at peak correlation plotted against QIDS difference, showing that E-field magnitude at this location was strongest for only a small subset of patients. C Thresholded E-field overlap across all patients, showing peak overlap at MNI coordinates [58,2,30] and [−46,28,38] (red crosshair). D E-field magnitude at these locations plotted against QIDS difference, showing no significant association. E Uncorrected voxel-wise correlation map of whole-brain E-field magnitude and QIDS improvement, adjusting for covariates. The whole-brain peak was at MNI coordinates [6,14,–6] (red crosshair), but did not survive multiple comparisons correction. F E-field magnitude at this peak location plotted against QIDS difference. Partial r and p-value refer to the results of the partial correlation analysis controlling for covariates. magnE: magnitude of electric field; MNI: Montreal Neurological Institute; QIDS: Quick Inventory of Depressive Symptomatology.

Full size image

Across patients, E-field coverage was maximal (n = 70/97 patients) at two specific locations, in the left (MNI:[-46,28,38]) and right DLPFC (MNI: [58,2,30]). At these locations, there was no linear relationship between E-field magnitude and clinical improvement (r = 0.070 and r = 0.009, respectively, Fig. 2C, D).

These results were not related to E-field thresholding, as similar results were seen when repeating the analysis with an unthresholded E-field (pFWE > 0.378; cross-validation r = 0.111, p = 0.287). The strongest uncorrected association was in the right nucleus accumbens (r = −0.420, MNI: [6,14,–6], Fig. 2E, F). Results were unchanged when modeling the coil at different hypothetical positions (Table S2).

Correlation between E-field connectivity and outcome

We tested whether E-fields improving depression are functionally connected to a distinct brain circuit. The strongest correlation was observed in the right middle frontal gyrus (r = −0.514, MNI: (46,8,52), pFWE < 0.001), indicating greater improvement for E-fields that are anti-correlated with this location (Fig. 3A, B). This circuit map will henceforth be described as the H1 coil depression circuit. Additional significant clusters included the caudate, right inferior parietal gyrus, right superior frontal gyrus, and the right periventricular hypothalamic nucleus (Table S3). Leave-one-out correlation maps were significantly associated with clinical improvement (r = 0.348, p = 0.0005).

Fig. 3: H1 E-fields that preferentially improve depression are connected to a distinct brain circuit.
figure 3

A Uncorrected voxel-wise correlation map of QIDS improvement and E-field connectivity modelled at 50 mm (top) and 45 mm (bottom) anterior to M1, controlling for age, sex and baseline symptoms. The two maps are significantly similar to each other, and survive multiple comparisons correction. B Individual E-field connectivity values at the peak voxel coordinates (MNI50mm: [46,8,52]; MNI45mm: [2,−2,−14], red crosshair in A) are plotted against QIDS difference. Partial r and p-values refer to the results of the partial correlation analysis controlling for covariates. C Leave-one-out (left) and absolute peak correlation values (right) are reported across models at various coil positions (at 40, 45, 50, 55 and 60 mm anterior to M1 on the cortex). The spatial correlation coefficients between the H1 coil depression circuits at the different locations are shown in the table to the far right, highlighted in bold when statistically significant with permutation testing. MNI Montreal Neurological Institute, FC functional connectivity, LOO leave-one-out, QIDS Quick Inventory of Depressive Symptomatology.

Full size image

We repeated the analysis for alternative hypothetical locations, hypothesizing that E-field models at locations other than the actual target would yield weaker results. Although the target is estimated at 50 mm anterior to M1 on the cortex (≅6 cm anterior on the scalp), the coil is often shifted posteriorly for tolerability. E-field connectivity measured at 45 mm anterior to M1 (≅5.4 cm on the scalp) yielded the strongest cross-validation value (r = 0.401, p = 0.0001, Fig. 3C), with a peak in the right periventricular hypothalamus near the septal nuclei (r = 0.503, MNI:[2,−2,−14], pFWE < 0.001, Fig. 3A). Additional significant clusters are reported in Table S4. Overall, the H1 coil depression circuits across target locations were highly similar (spatial correlation r = 0.91–0.99, pFWE < 0.05, Fig. 3C), although correlation strength and cross-validation values progressively decreased as moving the target farther away from the actual DLPFC target (Fig. 3C). These alternative models are summarized in Table S5 and Table S6.

Connectivity with an A priori convergent depression circuit

We observed a significant spatial correlation between the a priori convergent depression circuit [16] and our de novo H1 coil depression circuit at both the 50 mm (r = 0.58, pFWE = 0.05) and 45 mm locations (r = 0.59, pFWE = 0.04) (Fig. 4A).

Fig. 4: The H1 coil depression circuit is significantly similar to the a priori convergent depression circuit from Siddiqi et al. (2021).
figure 4

A Spatial correlation between the H1 coil depression circuit at 50 mm and 45 mm locations and the convergent depression circuit, significant after permutation testing. B Partial correlation coefficients for the relationship between QIDS difference and individual E-field connectivity similarity to the convergent depression circuit, reported across simulated coil positions. Peak correlation is observed at the 45 mm location (purple outline). C Scatterplot of QIDS difference and E-field connectivity similarity to the convergent depression circuit at the 45 mm location. Partial r and p-value refer to the results of the partial correlation analysis controlling for age, sex and baseline QIDS. D Partial correlation coefficients for the relationship between change in individual BASIS-24 items and E-field connectivity similarity to the convergent depression circuit at the 45 mm location. FC functional connectivity, cor: correlation, QIDS Quick Inventory of Depressive Symptomatology.

Full size image

The similarity of each patient’s E-field connectivity map with the convergent depression circuit was significantly associated with QIDS improvement across coil locations (Fig. 4B). The peak association with clinical outcomes was observed when the target was assumed to be 45 mm anterior to M1 (r = 0.411, p < 0.001) (Fig. 4C). This finding was consistent across multiple E-field thresholds (Table S5). Thus, r2 = 17% of the variance in clinical outcomes was explained by the targeted circuit.

Patients showing high E-field connectivity similarity to the convergent depression circuit showed greater average QIDS improvement (mean 47.67% (±26.71%)) relative to those with low similarity (mean 21.72% (±23.61%)) (t = 5.07, p < 0.001). This association was specific to depression relative to the individual items on the BASIS-24 (Fig. 4D).

Finally, in post-hoc analyses, we identified an optimized target maximizing the overlap of the H1 E-field with the convergent depression circuit. The optimal target was localized at MNI coordinates [−42,14,65], corresponding to a scalp location 3.6 cm anterior to M1. Overlap with the depression circuit progressively decreased with more anterior coil positioning, but remained strongly positive up to 6 cm anterior to M1 (Fig. 5).

Fig. 5: Simulations of optimal target location maximizing overlap with the convergent depression circuit.
figure 5

A Overlap between H1 E-field (thresholded to include the top 1% of voxels) with the convergent depression circuit, evaluated at different simulated coil locations. The coil center was systematically shifted in 5 mm increments along anterior (25–55 mm), lateral (20 mm) and medial (20 mm) directions from the MNI motor hotspot coordinate [−42,−16,65]. B Simulated target locations for the currently used target, the optimal target with maximal overlap, and the boundary delineating locations of lower overlap are projected onto the scalp using SimNIBS. MNI Montreal Neurological Institute.

Full size image

Discussion

Our findings indicate that effective H1 coil stimulation sites are functionally linked to a specific brain circuit. Notably, the H1 coil-specific circuit that we identified de novo is significantly similar to an a priori causal depression circuit derived from convergent data across lesions that cause depression, figure-of-8 coil TMS and DBS sites that modify depression [16]. Volumetric change in this circuit is also correlated with effectiveness of electroconvulsive therapy for depression [17]. This suggests that H1 coil TMS, despite its added depth and breadth, could be optimized by targeting the same circuitry as other forms of brain stimulation.

These findings align with a growing body of evidence emphasizing the importance of targeting brain circuits rather than individual regions [7, 13,14,15,16,17]. While this principle has been demonstrated in multiple studies using figure-of-8 coils, it remained unclear if targeting is relevant to clinical outcomes with broader and deeper coils. To our knowledge, this is the first study to demonstrate that the clinical efficacy of deep TMS is similarly linked to a brain circuit. Importantly, we derived this circuit in a data-driven manner and compared it to the a priori convergent depression circuit, without restricting our primary analyses to predefined circuits.

Patients whose E-field connectivity showed greater similarity to the convergent depression circuit experienced greater depression improvement. Despite inherent noise and imprecision in this analysis, similarity to this circuit explained 17% of the variance in clinical outcomes, and patients with above-median similarity to the circuit showed over twice as much improvement in depression relative to those with below-median similarity. This association was specific to depression relative to other behavioral domains. Although the derived H1-specific circuits were similar across hypothetical coil positions, the clinical effect of similarity to the a priori circuit was weakened when moving farther away from the actual DLPFC target. Overall, these findings suggest that, while the broader and deeper H1 E-field overlapped with the depression circuit in most cases, adjusting the coil position with respect to the network might further improve outcomes in some patients.

In post-hoc analyses, we identified an “ideal” stimulation site that maximizes overlap with the convergent depression circuit. Connectivity-based studies using figure-of-8 coils have generally suggested that stimulation may be more effective at more anterior and lateral DLPFC sites than those typically targeted using scalp measurements [32]. However, this has not been studied with the H1 coil, which has substantially different geometry relative to the figure-of-8 coil. Our results indicate that, with the H1 coil, a more posterior placement might be preferable, with a peak overlap observed at a scalp location approximately 3.6 cm anterior to M1. That said, targeting more posterior regions could increase seizure risk due to closer proximity to the motor cortex [22]. Of note, stimulation at 4 cm anterior to M1 appears to be safe with the H7 coil [33] and, according to our model, positioning the H1 coil up to 6 cm anterior to M1 would not substantially reduce overlap with the convergent depression circuit. Thus, we suggest that future prospective studies should investigate the safety, feasibility and efficacy of positioning an H1 coil less than 6 cm anterior to M1, as posteriorly as tolerable.

Strengths of this study include rigorous permutation-based statistics and alignment between data-driven and a priori hypothesis-driven results. There are also several limitations. First, this retrospective study utilized data from a patient database in a naturalistic treatment setting. While our findings align with previous literature, prospective validation of the proposed optimized target is essential. Second, we used group-based anatomical estimates of M1 and DLPFC coordinates, as individualized functional coordinates of motor hotspot or final coil position were not recorded. In addition, structural MRI scans used for E-field modeling were acquired at variable intervals from treatment. While E-field properties are unlikely to change substantially over time, this delay introduces potential measurement noise. Together, these factors may have reduced the precision of our E-field estimates and limited the observable effect size. Yet, we find that even with a crude estimate of coil placement, E-field connectivity to a depression circuit seems to impact clinical outcomes. Future studies that precisely record the coil location are necessary to quantify the effect size of targeting with an H1 coil. Third, we relied on normative connectivity data. Evidence from the DBS literature suggests that patient-specific structural connectivity and normative connectomes yield similar conclusions regarding the brain areas associated with clinical improvement [34]. Similarly, there is evidence suggesting that normative and disease-specific connectomes yield similar results [14, 35]. However, it is possible that individualized functional connectivity explains additional variance in outcomes [36]. Finally, our inclusion criteria might have introduced a bias towards more severe cases: the availability of a structural MRI in the patient’s record might indicate a more complicated medical history. Compared to the most recent post-marketing analyses on deep TMS efficacy [4], our sample indeed exhibited higher baseline severity and lower response and remission rates. Although QIDS-SR scores might be less sensitive to TMS-induced clinical changes compared to other clinical scales such as the PHQ-9 [37], this nonetheless raises questions about the generalizability of our findings, and highlights the need for replication studies.

Overall, similarity with an a priori depression circuit explained significant variance in outcomes, suggesting that targeting specific depression circuitry might be possible with an H1 TMS coil, despite its depth and breadth. Targeting this circuitry might be optimized with more posterior coil placement, but prospective studies are needed to evaluate the safety, feasibility and efficacy of such adjustments to H1 coil positioning.