Relating Symptom Clusters to Neuroimaging Findings: New Analysis From GEMRIC

Out on PubMed, from an international group of investigators, is this article:

Accounting for symptom heterogeneity can improve neuroimaging models of antidepressant response after electroconvulsive therapy.

Wade BSC, Hellemann G, Espinoza RT, Woods RP, Joshi SH, Redlich R, Dannlowski U, Jorgensen A, Abbott CC, Oltedal L, Narr KL.Hum Brain Mapp. 2021 Aug 13. doi: 10.1002/hbm.25620. Online ahead of print.PMID: 34390089

The abstract is copied below:

Depression symptom heterogeneity limits the identifiability of treatment-response biomarkers. Whether improvement along dimensions of depressive symptoms relates to separable neural networks remains poorly understood. We build on work describing three latent symptom dimensions within the 17-item Hamilton Depression Rating Scale (HDRS) and use data-driven methods to relate multivariate patterns of patient clinical, demographic, and brain structural changes over electroconvulsive therapy (ECT) to dimensional changes in depressive symptoms. We included 110 ECT patients from Global ECT-MRI Research Collaboration (GEMRIC) sites who underwent structural MRI and HDRS assessments before and after treatment. Cross validated random forest regression models predicted change along symptom dimensions. HDRS symptoms clustered into dimensions of somatic disturbances (SoD), core mood and anhedonia (CMA), and insomnia. The coefficient of determination between predicted and actual changes were 22%, 39%, and 39% (all p < .01) for SoD, CMA, and insomnia, respectively. CMA and insomnia change were predicted more accurately than HDRS-6 and HDRS-17 changes (p < .05). Pretreatment symptoms, body-mass index, and age were important predictors. Important imaging predictors included the right transverse temporal gyrus and left frontal pole for the SoD dimension; right transverse temporal gyrus and right rostral middle frontal gyrus for the CMA dimension; and right superior parietal lobule and left accumbens for the insomnia dimension. Our findings support that recovery along depressive symptom dimensions is predicted more accurately than HDRS total scores and are related to unique and overlapping patterns of clinical and demographic data and volumetric changes in brain regions related to depression and near ECT electrodes.

Keywords: electroconvulsive therapy; machine learning; major depressive disorder; structural neuroimaging; symptom heterogeneity.

From the Introduction:

More recently, we used EFA to identify latent symptom dimensions in a large, multisite patient cohort undergoing electroconvulsive therapy (ECT) for depression (Wade et al., 2020). We identified three symptom dimensions from pretreatment HDRS-17 items: somatic symptoms, core mood and anhedonia, and insomnia. Here, we build on this work and aim to relate patient demographic, clinical, and patterns of regional ECT-induced volumetric brain changes to changes in latent symptom dimensions using data-driven methods. We hypothesized that demographic and clinical measures would be broadly related to symptom changes, and as has been previously reported, that regional volumetric changes of the hippocampus, motor cortex, striatum, and cingulate would be commonly related to changes along separate symptom dimensions. However, at the same time we hypothesized that there would be unique volumetric changes related to individual symptom dimensions based on the assumption that these dimensions recruit different functional networks. Finally, we hypothesized that prediction of symptom changes would be more accurate along latent symptom dimensions rather than the HDRS-17 total score. Characterizing these unique and shared predictors and mechanisms of treatment response will inform development of targeted interventions such as neurostimulation techniques capable of targeting specific neural circuits and thus directly targeting important dimensions of depression.

From the Discussion:

Our study extends recent work using exploratory factor analysis to identify three homogenous latent symptom dimensions of treatment-resistant MDD. We used data-driven methods to identify clinical, demographic, and distributed patterns of regional volumetric brain changes associated with ECT-related recovery along these latent symptom dimensions of depression. Models predicting recovery along core mood and anhedonia and insomnia dimensions were predicted significantly more accurately than the more widely used HDRS-6 and HDRS-17 total scores. Further, treatment-related recovery along these symptom dimensions was associated with unique and overlapping sets of clinical, demographic, and volumetric brain regions. Brain regions informative of symptom changes were generally well-aligned with existing literature.

Easily-acquired and inexpensive clinical and demographic measures including pretreatment symptom severity, patient age, and body mass index (BMI) were generally the most informative predictors of change. Previous studies have reported that increased symptom severity and presence of psychotic features are indicative of better response to ECT (van Diermen et al., 2018). Increased patient age has widely been linked with better ECT responsivity (Kranaster et al., 2018; O'Connor et al., 2001). However, the etiology of depression in elderly patients more frequently involves cardiovascular pathology, cognitive impairment, or chronic medical illness (Alexopoulos, 2005), which in addition to normal aging effects of brain tissue loss, could impact response to ECT. In our sample, patient age differed significantly by site that may have limited our model generalizability to new sites. Earlier work has also reported that elevated BMI is associated with reduced white matter integrity (Repple et al., 2018), gray matter reductions of frontal regions, and a more chronic course of depression (Opel et al., 2015). Interestingly, while elevated BMI predicted poorer outcomes for the CMA symptom dimension, it predicted better treatment response for the SoD and insomnia symptom clusters. This effect is likely driven by the inclusion of the HDRS weight loss item in the CMA factor. Here, we observed that, relative to patients who exhibited no change in the weight loss item, those who reduced their weight loss symptoms had a significantly lower pretreatment BMI (p < .05).

From Limitations:

Naturalistic treatment protocols across this multisite study are a potential limitation; however, this design also builds the natural diversity of ECT treatment protocols and outcomes seen worldwide into our models. Variation in treatment parameters was naturally confounded with acquisition site. We did, however, include both site and electrode placement as predictors though electrode placement was not an important predictor. Relatedly, patient inclusion/exclusion criteria varied by site though all had in common that patients were required to have failed at least one prior treatment or that patients immediately needed ECT. A minority of patients were included with diagnoses of bipolar disorder and depression with psychotic features. Symptom changes did not significantly differ among those with bipolar disorder but were significantly more reduced among those with psychotic features as is commonly observed (van Diermen et al., 2018). Site 1 patients were also least responsive with an average HDRS-17 reduction of 3.8 points. These patients were tapered off of antidepressant medication prior to ECT index, though concurrent medications do not reliably affect the efficacy of ECT (Fink, 1994; Haskett & Loo, 2010). Exclusion of patients from Site 1 largely improved model performance and generalizability but this gain is counterbalanced against the benefit of training these models using the range of ECT outcomes seen around the world. Lastly, we emphasize that prospective prediction of clinical outcomes was not attempted; this would require exclusive use of pretreatment predictors.

 From Conclusions and future directions:

Taken together, our findings provide new evidence that use of homogenized latent symptom dimensions of multi-item scales can improve the detection of imaging, demographic, and clinical biomarkers related to the trajectories of specific symptom constellations. While clinical and demographic measures accounted for more outcome variability, neuroimaging measures of regions often implicated in the pathology of depression and ECT-related treatment response were significantly predictive and accounted for between 1 and 10% of outcome variability when used alone. As neurostimulation methods become more refined and capable of targeting more specific neural systems, it is plausible that findings such as these will inform the targeting of neural systems underlying more specific symptom dimensions. Future work will explore whether prospective prediction of change in these symptom dimensions will similarly be predicted more accurately than more heterogeneous total scores.

This is a remarkably sophisticated and dense paper. The idea behind it is simple: are there symptoms, or symptom clusters that correlate with improvement and brain structural changes on MRI? The statistical and imaging methods are highly advanced. The results are complex. The figures are gorgeous.

But there are so many confounds in the data (lack of standardized ECT protocols, with different electrode placements, different stimulus dosing methods, different ways of determining clinical endpoints, etc.) that one must ask if the silk purse of the data analysis is a real Gucci, or a knockoff?

In other words, can sophisticated science and math overcome the heterogeneity of the clinical data and the reality of crude data collection methods? There is irony here: the overarching effort is to improve psychiatric diagnosis and prognostication, but the inescapable, old inputs suffer from the inclusiveness of DSM ("major depression"), the shortcomings of the HAM-D, and reliance on the training of the interviewer, even if the outputs look cutting edge. IMO, the experienced clinician has not yet been replaced by exploratory factor analysis...

As I have discussed recently, the effort to focalize ECT along the lines of other brain stimulation techniques, ignores the value of the "broad spectrum" effects of ECT against a variety of severe psychiatric illnesses.

I recommend that blog followers read this article in full (~30-45 mins.) and comment with your opinions, please. Kudos to the UCLA and GEMRIC teams for bringing us this thought-provoking work.