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

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 significance [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 findings further confirm this relationship in an array of groups with different and often mixed electrode placements (RUL, BT, and MIX). The previous findings 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 confirms that EF modeling, despite its limitation [34], is a useful technique to estimate EF.

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 CDN 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 confidence about the validity of these findings. First, it was unsupervised and data-driven, to avoid overfit and no information about the CDN 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 et al. As reported in patients treated with DBS and TMS, individual similarity to this map correlated with antidepressant outcome in ECT.

A set of interconnected areas, with sometime opposite signs in their relationship, is reliably implicated in association with depression. In addition to the original discovery, a very recent lesion mapping study in patients with multiple sclerosis resulted in a very similar map, showing close correlation with depression rates. While our methodology deviates from traditional lesion mapping, as it is measuring structural volume change patterns, the second principal component of the variance shows a distinct similarity with the networks reported previously. The brain wide volume changes which failed to associate with clinical response likely includes both epiphenomena and antidepressant volumetric changes [17] Our results suggest that the antidepressant volumetric changes are relatively hidden (PC2_ΔVOL) behind the main volumetric effect of ECT (PC1 _ΔVOL). This explains why previous univariate approaches failed to detect correlations between volume change and clinical response. Without decomposing the volume changes to a main effect (PC1 _ΔVOL), which is responsible for most of the variance, and to an orthogonal pattern (PC2 _ΔVOL), this second pattern would remain hidden. Our results indicate that only this second pattern of change, which is similar to the pattern obtained by Siddiqi et al, has clinical relevance. The second component explains a relatively small amount of variance (6–11%) compared to the first component (~40%). This raises the question of what, if anything, PC1 _ΔVOL is associated with. Earlier findings also suggest that certain volume increases can lead to cognitive side effects often linked with ECT [35]. Unfortunately, we were unable to examine whether PC1_ΔVOL and PC2_ΔVOL can differentiate these effects due to the unavailability of harmonized neurocognitive data across different sites. The overall interpretation of these two volumetric PCs was reinforced by the observed correlations between their values and the number of ECT sessions. The positive relationship between PC1_ΔVOL and the number of ECT sessions suggests a dose response relationship that is independent of the clinical response. Although this correlation is modest, it is important to acknowledge that the number of ECT sessions serves as a rudimentary approximation of the overall ECT dosage, which can vary individually based on the seizure threshold. On the contrary, the inverse correlation between PC2_ΔVOL and the number of ECT sessions arises from an indirect association, given that a higher number of ECT sessions was linked with poorer outcomes in this observational cohort of patients (as outlined in the methodological considerations above).

Considering the convergence of findings between TMS and DBS-based CDN, along with our PC2_ΔVOL, it prompts the question of whether there exists a volumetric effect of DBS and TMS. Although human studies reporting structural changes after DBS in depression are limited [36], animal studies have demonstrated that deep brain stimulation of the ventromedial prefrontal cortex leads to a distant effect, resulting in increased hippocampal and thalamic volumes [37]. In the case of TMS, there is more evidence for macroscopically measurable effects across the brain. It has been demonstrated that as few as five TMS treatments can lead to measurable macroscopic changes in the superior temporal cortex [38]. Subsequent studies [39, 40] implicated the rACC not only in volume changes but also in its correlation with clinical outcomes. The rACC is one of the peak regions in the PC2 and CDN, exhibiting a notable convergence with both this region and the prior literature [41]. It is important to acknowledge a noteworthy contradiction here, as our results show the opposite sign compared to previous TMS studies. However, our results indicate a “residual” effect over the main effect, which is an overall increase in ECT. Therefore, careful interpretation is needed when comparing these results with distinct methodologies.

Furthermore, multivariate analysis of the EF revealed two components: the first represented the main effect, the overall strength of the EF across the brain, and the second was specific to the spatial particularities of the electrode placement. The first component that reflected the overall EF strength correlated with clinical response, indicating that higher EF was associated with worse outcome. We also found that the higher the first component of EF was expressed, the lower the CDN expression in the patient (the second component of the volume change). In a classical sense [42] 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 beneficial 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 first, but supported by one previous study in a smaller, and only in BT treated cohort [21], and might indicate that more focal or low amplitude treatments would be more beneficial. This seemingly contradicts some of the clinical observations in recent studies of RUL patients where the amplitude of the current was modified and was found that below certain range the clinical effect was insufficient [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) [43]. 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 first PC, representing the main effect, was similar regardless of the analytical approach (whole brain, right or left side, Supplementary Fig. 9). There were, however, hemispheric differences in the second PC_ΔVOLs: 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 [44, 45].

The study has some noteworthy limitations. First there are significant methodological differences between the original study and the current approach. While the original approach was a voxel based lesion network mapping, this is ROI based multivariate structural analysis. It’s worth noting the potential drawback to this method is that certain aspects of the topography in the Siddiqi et al. map may be compromised when transformed into larger ROIs covering multiple functional regions. Secondly, our analysis was limited to identifying associations between imaging results and depression outcomes. Therefore, we were unable to fully demonstrate the specificity of our findings or to test other associations. Thirdly, we recognize that it would have been ideal to model not only the E-field but also the ECT dose. Unfortunately, only the number of ECT sessions was available for this cohort, serving as a crude proxy for estimating the overall ECT dose. Lastly, the variability in potential image quality and clinical practices among different sites could have introduced some level of noise into the analysis. It was evident that certain sites predominantly utilized either RUL or BT, thereby introducing a potential bias in electrode placement across the various sites.

In summary, this current report is a significant step forward in understanding the direct electrical stimulation aspect of ECT and its effect on brain volume changes and clinical outcomes. However, our findings, in their present state, do not offer direct clinical guidance, as we have yet to fully characterize the relationship between ECT dosage, electrode placement, network formation timeline, and specificity.

To address this gap, we must investigate the dose-response dynamics of volume change pattern formation, considering factors such as (1) EF strength, (2) electrode placement, (3) dosage, and (4) seizure patterns, all of which influence volume changes. This understanding is crucial for designing prospective clinical studies that aim for optimal benefit. The studies require ECT machines capable of varying current amplitude, random electrode placement assignment to mitigate site-specific effects, and ideally a consistent number of stimuli. Our deepening grasp of this domain can guide the development of well-informed studies and methodologies to refine treatments.

To establish causality, we can test if parameter settings that enhance CDN formation in volume changes lead to improved treatment effects. Concurrently, future ECT studies should explore similar networks in psychotic disorders to discern whether a transdiagnostic signature exists or if patterns differ across disorders.

Overall, the revelation that the same neural network associated with clinical benefits in TMS and DBS is also implicated in ECT offers promise. It suggests that delving further into ECT-related clinical networks could shed light on these other treatment modalities as well.