Electroconvulsive therapy-induced volumetric brain changes converge on a common causal circuit in depression

Neurostimulation is a mainstream treatment option for major depression. Neuromodulation techniques apply repetitive magnetic or electrical stimulation to some neural target but significantly differ in their invasiveness, spatial selectivity, mechanism of action, and efficacy. Despite these differences, recent analyses of transcranial magnetic stimulation (TMS) and deep brain stimulation (DBS)-treated individuals converged on a common neural network that might have a causal role in treatment response. We set out to investigate if the neuronal underpinnings of electroconvulsive therapy (ECT) are similarly associated with this common causal network (CCN). Our aim here is to provide a comprehensive analysis in three cohorts of patients segregated by electrode placement (N = 246 with right unilateral, 79 with bitemporal, and 61 with mixed) who underwent ECT. We conducted a data-driven, unsupervised multivariate neuroimaging analysis (Principal Component Analysis, PCA) of the cortical and subcortical volume changes and electric field (EF) distribution to explore changes within the CCN associated with antidepressant outcomes. Despite the different treatment modalities (ECT vs TMS and DBS) and methodological approaches (structural vs functional networks), we found a highly similar pattern of change within the CCN in the three cohorts of patients (spatial similarity across 85 regions: r = 0.65, 0.58, 0.40, df = 83). Most importantly, the expression of this pattern correlated with clinical outcomes. This evidence further supports that treatment interventions converge on a CCN in depression. Optimizing modulation of this network could serve to improve the outcome of neurostimulation in depression.


Introduction
One of the oldest and most effective forms of neurostimulation is electroconvulsive therapy (ECT) 1,2 However, despite the last decades of ECT-neuroimaging research, its mechanism of action is not known.
In a recently published article, Siddiqi et al. 3 showed that a common underlying neural network ( Supplementary Fig. 1A) is associated with the clinical response of treatment resistant depression in transcranial magnetic stimulation (TMS) and deep brain stimulation (DBS). Additionally, dysfunctions within this network explain clinical symptoms in patients with stroke, multiple sclerosis and other forms of brain lesions 3,4 . These results are not mere associations, but instead indicate that interference within this network could explain individual differences in treatment response. The main cortical areas associated with the common causal network included regions previously implicated in depression and emotion regulation, such as the subgenual cingulate cortex, dorsolateral prefrontal cortex, ventromedial prefrontal cortex, inferior frontal gyrus, frontal eye eld, and intraparietal sulcus ( Supplementary Fig. 1A, 3,5 ).
While ECT is not a localized form of treatment, the applied ECT-induced electric eld (EF) has unique spatial distribution speci c to an individual and the electrode placement 6-9 . High frequency electric eld (EF) stimulation has a direct neuroplastic effect on the brain [10][11][12] and is also associated with downstream biological effects through the induced seizure activity 13,14 . In agreement with these preclinical ndings, recent large-scale studies in the Global ECT-MRI Research Collaboration (GEMRIC) dataset 15 found robust volume increases 16,17 in a wide range of cortical and subcortical regions, which correlated with the number of ECT sessions. Subsequent EF modeling based on the individual head MRI consistently demonstrated that the ECT-induced EF strongly correlated with volume increase 18-20 . These results veri ed that despite the widespread activation of the brain through seizure, the direct electrical stimulation effect of ECT is much more spatially selective and individually diverse than rst assumed.
Despite these replicable and robust structural ndings driven by the spatial distribution of the EF, their direct or indirect effect on clinical outcome remain unclear. Univariate analysis of the EF amplitude on clinical response was the subject of several previous investigations. However, the results were somewhat contradictory 18, 21,22 . Similarly, the robust changes in volume did not translate into correlations with clinical response in most of the studies 16,17 or indicated a relationship where volume increase in the dentate gyrus was associated with worse clinical outcome 23 .
One caveat was the primarily univariate nature of these analyses. The brain regions are not independent of each other, and multivariate analysis could be more sensitive to detect network-wide changes 3 . Indeed, one follow-up analysis of the GEMRIC dataset in 192 individuals with supervised multivariate models could detect networks of regions where the weighted average of the changes correlated with clinical outcomes 24 . The results of this analysis showed that the linear combination of volume changes across 18 regions correlated with clinical outcomes. The loadings of the 18 regions showed some similarities with the spatial distribution of the causal map published by Siddiqi et al. (Supplementary Fig. 1B, r = 0.33, df = 16). Although these 18 regions comprise a limited coverage of the causal map, it raises the intriguing possibility that the ECT-induced volume changes might follow a similar spatial pattern already described with functional connectivity analysis in other treatment modalities such as TMS and DBS.
To address this question, we revisit and improve the analysis of EF-structure to clinical outcomes by doubling the sample size across three independent cohorts (total N = 386). We implement an unsupervised learning algorithm (Principal Component Analysis, PCA) running separately on EF and structural data, and separately on different electrode placements (6 parallel PCAs). The non-supervised learning methods reduce the risk of over tting. Any convergence in these independent but parallel multivariate analyses would strongly support the validity of our ndings and indicate a common pathway in the mechanism of action of ECT. Finally, we will demonstrate not only that a common circuit emerges, but that it is remarkably similar to the causal circuit previously published in the context of TMS and DBS e cacy 3 .

Methods
Participants 386 ECT-treated subjects were analyzed from the GEMRIC consortium 15 . This multi-site consortium collects data in a centralized server from ECT-treated patients who underwent longitudinal neuroimaging and clinical assessment. The 386 subjects were recruited at 19 sites and their respective demographics and clinical data are in Supplementary Table 1. All contributing sites received ethics approval from their local ethics committee or institutional review board. In addition, the centralized mega-analysis was approved by the Regional Ethics Committee South-East in Norway (No. 2013/1032).

Volume changes
The image processing methods have been detailed previously [16][17][18] . In brief, the sites provided longitudinal 3T T1-weighted MRI images (at baseline and after the end of the course of ECT) with a minimal resolution of 1.3 mm in any direction. The raw DICOM images were uploaded and analyzed on a common server at the University of Bergen, Norway. To guarantee reproducibility, in addition to the common platform, the processing pipelines were implemented in a docker environment 25 . First, images were corrected for scanner-speci c gradient-nonlinearity 26 . Further processing was performed with  Fig. 3). In more detail, we cross-sectionally processed both time points separately with the default FreeSurfer work ow and created an unbiased template from both time points for each subject. Once this template is created, parcellations and segmentation are carried out at each time point initialized with common information from the within-subject template 29 . In summary, we calculated biasfree estimation of volumetric change from 85 brain regions across the timespan of an ECT course in 386 individuals who received on average of 12.5 ± 5.4 ECT sessions.

EF modeling
Our approach was detailed in one of our previous manuscripts 18 , with the upgraded software of Roast 3.0 (Realistic Volumetric-Approach to Stimulate Transcranial Electric Stimulation v3.0) 6 . In short, ROAST builds a three-dimensional tetrahedral mesh model of the head based on the T1 MRI images of the brain.
Then, segmentation identi es ve tissue types: white and gray matter of the brain, cerebrospinal uid, skull, and scalp, and assigns them different conductivity values: 0.126 S/m, 0.276 S/m, 1.65 S/m, 0.01 S/m, and 0.465 S/m respectively. ECT electrodes of 5 cm diameter were placed over the C2 and FT8 EEG (10-20 system) sites to model RUL, and over to FT8 and FT9 sited to model BT electrode placements. Study sites from the GEMRIC database used either the Thymatron (Somatics, Venice, Florida) or spECTrum (MECTA Corp., Tualatin, Oregon) devices. EF was solved using the nite-element method with unit current on the electrodes and, subsequently, it was scaled to the current amplitude of the speci c devices (Thymatron 900 mA, spECTrum 800 mA). We had 61 individuals who had to switch from RUL to BT electrode placement during the ECT course. This is a standard clinical practice in patients with inadequate clinical response with RUL stimulation. In these cases, we calculated the EF with the weighted mean according to the number of ECT sessions the individual had in each form of placements. For example, if a patient had 6 ECTs with RUL and then had 18 ECTs with BT then we calculated 0.25 x EF RUL + 0.75 x EF BT in each region. These procedures resulted in a voxel-wise EF distribution map in each individual. We calculated the average EF across the 85 three-dimensional ROIs at baseline in every individual based on the Freesurfer parcellations and segmentations. The voxel values with the top and lowest one percentile in each ROI were omitted during calculations to reduce boundary effects.

Multivariate analysis
To investigate the regional volume changes and EF amplitudes in a multivariate way, we applied principal component analysis (PCA). We conducted six consecutive PCA analyses on RUL, BT, and MIX separately for EF and structural data, respectively (variables were normalized across individuals before PCA). We separated the groups as we wanted to avoid capturing differences that were only electrode placement speci c. We used Cattell's scree test to determine the number of PCs to analyze. We found that the rst 2 PCs captured most of the variance, and the subsequent PCs captured a diminishing portion of the variance (elbow criteria, Supplementary Fig. 2). We conducted posthoc analyses to evaluate 1) the correlation between PCs and clinical outcomes: ΔMADRS ~ PC1 + PC2 + age + nECT (nECT: number of ECT sessions, ΔMADRS: percent change compared to baseline (T2-T1)/T1, negative values indicate better response), and 2) the spatial similarity between loadings and the CCN 3 . To investigate if one hemisphere was driving the results, we conducted the PCA separately on the right and left hemisphere (Supplementary Material).

Covariates
We conducted multivariable regression analysis to estimate the effect of the calculated principal components on clinical response. This analysis included the principal components of the volume change, EF, age and number of ECT sessions as independent variables. These last two variables were included as confounders. As it is explained below, age correlated with EF and clinical response, and number of ECT sessions were also correlating with clinical response and volume change. Therefore, these variables had to be added to correct for spurious correlations.

Justi cation of the confounding variables
We corrected for two variables consistently across our analyses. We would like to provide a brief justi cation for including these. We also provide a causal model with a corresponding directed acyclic causal graph to illustrate the reasoning (Supplementary Fig. 7).

Number of ECT sessions
It was already noted in the rst large scale publication of the GEMRIC consortium that the number of ECT sessions and clinical response correlated in a counterintuitive way: the larger the number of ECT sessions registered between MRI assessments, the lower the clinical response was. The explanation of this observation is that most of the participating sites in the GEMRIC consortium acquired the follow-up (post-ECT) MRI after completing the (un-) successful ECT course, in contrast to predetermined length of treatment period with a xed number of ECT-sessions. This resulted in an earlier timepoint of post-ECT MRI assessment if there was a quick clinical response, but later when clinical improvement was delayed or absent. This is problematic because the number of ECT sessions positively correlates with the volume change during ECT (dose-response effect). Therefore, not controlling for the number of ECT sessions can easily lead to spurious correlations indicating that volume increase was associated with worse outcome, or just simply mask the otherwise real effect when volume change is bene cial. Indeed, in recent cohorts where the length of ECT course between the neuroimaging sessions were predetermined, authors found positive relationships between hippocampus volume increase and clinical response 19 ( Table 1).

Age
Our sample had a tight correlation between age and clinical response as well. This correlation is typical in ECT datasets [30][31][32] , as the elderly patients respond to ECT signi cantly better. This introduces, however, another confound to every EF modeling as age negatively correlates with EF magnitude in the human brain due to structural changes such as atrophy 9 . This age and EF relationship was particularly strong in RUL (right unilateral) placement (R Hippocampus; RUL: r=-0.31, p < 0.001, df = 244, BT: r=-0.17, p = 0.13, df = 77, MIX: r=-0.28, p = 0.03, df = 59), therefore it could mask the effect of EF on clinical response in our previous analysis 18 .
The code relevant to this manuscript is available at https://github.com/argyelan/Publications/tree/master/VOLUME-CHANGE-PCA .

EF correlates with volume changes
The dataset is detailed in Supplementary Table 1A and B. Like our previous studies, we found volume increases in almost every region across the brain with effect sizes (Cohen's d) ranging from − 0.02 to 1.93, corresponding 0% (Left Cerebellum) to 6.7% (Right Amygdala) volume increases. 75% of the 85 regions had a volume increase of at least 0.5 effect size or higher (t > 9.8, p < 10 − 12 , df = 385, Supplementary Table 2A). 246 patients received right unilateral (RUL) ECT placement only, 79 bitemporal (BT) only, and 61 individuals rst started with RUL and were later switched to BT (MIX) treatment. The overall volume changes were higher in the BT and MIX groups than in the RUL group (mean volume increase: 3.5% ± 1.7%, 3.4% ± 2.1% vs 1.7% ± 1.  Table 3B).

Unsupervised multivariate analysis
Volume changes: In agreement with previous ndings 17 the rst PC (PC1) ( Fig. 2A left) was responsible for 42%, 42%, and 41% of the variance in the volume changes in the RUL, BT, and MIX groups, respectively. This 42% variance indicated a strong intra-individual cross-correlation in regional volume increase. The loadings of this main effect showed spatial similarity with the common causal network (RUL: r = 0.44, p = 2x10 − 5 ; BT: r = 0.50, p = 1x10 − 6 , MIX: r = 0.46, p = 8x10 − 6 ; df = 83) even though it was an unsupervised nding. The second PC (PC2) ( Fig. 2A right) was responsible for 6%, 8%, and 11% of the variance and the loading was spatially very similar to the common causal network (RUL: r = 0.65, p = 2x10 − 11 ; BT: r = 0.58, p = 6x10 − 9 , MIX: r = 0.40, p = 0.0002; df = 83, Fig. 2C). Moreover, the effect sizes of similarity of loadings of the second component were higher in all groups than in the rst component.

Spatial coordinates as covariates
We conducted a multiple regression of the "CCN values" ~ abs(X) + Y + Z, where X, Y, and Z were the coordinates across the LR (left-right), PA (posterior-anterior) and the IS (inferior-superior) axes, respectively. The Siddiqi map values showed strong correlations across the spatial dimensions of the regions, especially across the posterior-anterior and inferior-superior axis (F 3,81 = 19.6, p = 1x10 − 9 ; t Xabs = 5.1, p < 0.0001, t Y = -2.2, p = 0.03, t Z = 5.2, p < 0.0001), indicating higher values on the lateral and on the superior areas. The solid spatial similarities between the CCN and the main (loadings of PC1) and the secondary effect (loadings of PC2) raised the question if these maps were only re ecting gross similarities across the posterior-anterior or inferior-superior direction. One could argue that both RUL and BT had a higher impact on the superior and lateral areas, resulting in a more reliable volume change in these regions leading to a PC that had grossly matching spatial distribution with the CCN. We tested this hypothesis by conducting a multiple regression with spatial coordinates of the regions as confounders: "CCN values" ~ PC1 + PC2 + X + Y + Z in all three groups (RUL: In both RUL and BT groups, the results equivocally identi ed that PC2 showed highly signi cant similarities that could not be explained by gross anatomical similarities (Supplementary Table 4A, B, and C).
EF amplitude: The rst PC (Fig. 2B left) was responsible for 70%, 65%, and 57% of the variance in the EF amplitude in the RUL, BT and MIX groups, respectively. This high variance in the rst PC indicated a strong intra-individual correlation across the brain regions, meaning that individuals with high EF had higher EFs across different regions. Therefore, this rst PC represented an overall EF magnitude across subjects, which was due to individual differences in brain and head anatomy, including the amount of cerebrospinal uid and fat tissue. The second PC (Fig. 2B right) was responsible for 7%, 10%, and 24% of the variance, respectively. The spatial distribution of the second PC re ected the electrode placement, showing higher loading near the electrode locations. The loadings of PC2 did not show any signi cant correlation with the CCN in RUL and BT. In the MIX group, the PCA analysis indicated that the main (PC1) and electrode effect (PC2) was more interleaved, re ecting in the lower and higher variances in the rst and second PC. This was also re ected in its loading structure. Overall, none of the PCs from the EF amplitudes showed any correlation with the CCN once it was corrected for the spatial coordinates (Supplementary Tables 4A, B

Laterality
Finally, we repeated the same multivariate analyses on the right and the left side of the brain independently (42 ROIs in each hemisphere). We found this necessary for two reasons. First, there is a large literature indicating differences among hemispheres in mood regulation 33 . Second, both the map reported by Siddiqi et al. and the bilateral placement maps are highly symmetrical which could lead to a spurious overestimation of the results.
The loadings of the PCs acquired on the entire brain appeared highly symmetrical ( Supplementary Fig. 4). As expected, the PCs of the EF in the RUL setting were the least symmetrical, but the volume changes were more symmetrical across different electrode placements. When we ran the PCAs separately for the right and left hemispheres (PCA right , PCA left ), we found that the results were very similar for the PCA right (Supplementary Figs. 5 and 6, r: 0.72-0.99). We found that the second PC of volume change highly correlated with the CCN and with the clinical response (Supplementary Material). The second PC of volume change also correlated highly with the loadings of the original PCA ( Supplementary Fig. 6). The PCA left led to a similar loading structure in their rst PCs of the EF and volume changes (main effect, r = 0.61-0.95), but the second PCs were different in RUL and BT (r = 0.01, 0.47, respectively) and volume change did not correlate with clinical outcome (Supplementary material).

Discussion
Our study is a comprehensive multivariate analysis of 386 patients with depression who underwent ECT and longitudinal neuroimaging. Our multivariate non-supervised analysis (PCA) revealed a hidden pattern in volume change that was correlated with clinical outcome. The same pattern was found independently in the three separate groups, RUL, BT and MIX electrode placements, and this pattern showed striking similarities to the common causal circuit recently published in a study of large cohort of independent samples of depressed patients 3,4 . This network consists of cortical and sub-cortical areas previously implicated in depression or emotion regulation, such as the subgenual cingulate cortex, dorsolateral prefrontal cortex, ventromedial prefrontal cortex and hippocampus.
Initial studies of ECT effect on structural neuroimaging on limited sample sizes (N ~ 20) often focused on hippocampus increase. As it became clear later, more widespread volume changes with moderate effect sizes were present in these samples, but due to the limited sample size, it did not reach statistical signi cance 15 . After establishing the GEMRIC consortium 15 and collecting hundreds of individuals, ECT studies repeatedly and consistently showed increased volume in both cortical and subcortical regions [16][17][18] . The GEMRIC data also demonstrated that the volume increase correlated with the EF amplitude in RUL electrode placement 18 . This relationship between EF and volume change was recently replicated in an independent cohort 19 . The current ndings further con rm this relationship in an array of groups with different and often mixed electrode placements (RUL, BT, and MIX). The previous ndings were replicated, and the weighted mixing of the EF values according to the number of ECT sessions on different electrode placements proved to be a useful way to calculate the effect of EF on volume change (MIX electrode placement). Our current study further con rms that EF modeling, despite its limitation 34 , is a useful technique to estimate EF, and the growing body of evidence of the correlation between EF and volume changes serves as biological validation of the technique.

Multivariate analysis and clinical effect
We found a spatial pattern in the volume changes on top of the main effect which showed distinct similarities to the CCN map reported by Siddiqi et al and was responsible for approximately 6-11% of the total variance. Most importantly, the more this pattern was expressed, the better the clinical outcome was. Our approach had two vital aspects that could further boost con dence about the validity of these ndings. First, it was unsupervised and data-driven, to avoid over t and no information about the CCN was used to conduct our analysis. Second, we analyzed the three electrode placement groups separately as independent samples. We not only received equivalent results across distinct groups, but these results were highly similar to the common causal circuit reported by Siddiqi  Furthermore, multivariate analysis of the electric eld revealed two components: the rst represented the main effect, the overall strength of the EF across the brain, and the second was speci c to the spatial particularities of the electrode placement. The rst component that re ected the overall EF strength correlated with clinical response, indicating that higher EF was associated with worse outcome. We also found that the higher the rst component of EF was expressed, the lower the CCN expression in the patient (the second component of the volume change). In a classical sense 35 our multiple regression analysis between clinical effect and the principal components of the EF and ΔVOL might imply that the high EF was mediated through a lower expression of the bene cial pattern, leading to a less than optimal outcome. As there are several limitations to deduct causal inference from classical mediation analysis, we point out here the need of prospective studies with preselected parameters to determine true cause and effect relationships in the future. The observation that high general EF was associated with worse clinical outcome is counterintuitive rst, but supported by one previous study in a smaller, and only in BT treated, cohort (Fridgeirsson et al. 2021), and might indicate that more focal or low amplitude treatments would be more bene cial. This seemingly contradicts some of the clinical observations in recent studies of RUL patients where the amplitude of the current was modi ed and was found that below certain range the clinical effect was insu cient 22 . However, these tested values (600, 700 and 800 mA) were below the range we use in current clinical practice and constitute the database analyzed in this article (800 and 900 mA). These results suggest that too low or too high EF might be equally suboptimal (inverted U hypothesis of strength of EF and antidepressant outcomes) 36 . The optimal strength of EF and its proper spatial distribution is an intriguing new direction that must be systematically investigated in the future.
Finally, PCA on each hemisphere independently showed that the rst PC, representing the main effect, was similar regardless of the analytical approach (whole brain, right or left side, Supplementary Fig. 6). There were, however, hemispheric differences in the second PCs: whole-brain PCA results were only replicated on the right side. This implies that neuroplastic changes associated with clinical outcomes were more robust on the right side 37,38 .
In summary, this current report is a signi cant step forward in understanding the direct electrical stimulation aspect of ECT and its effect on brain volume changes and clinical outcomes. Our deepening understanding in this domain could lead to informed decisions in designing novel studies and methods to optimize treatments. In addition, this study revealed that the same neural network associated with clinical bene ts in TMS and DBS is implicated in ECT as well.
Declarations Table   Table 1 is not available with this version.  Supplementary Files