A multivariate neuroimaging biomarker of individual outcome to transcranial magnetic stimulation in depression

Abstract

The neurobiology of major depressive disorder (MDD) remains incompletely understood, and many individuals fail to respond to standard treatments. Repetitive transcranial magnetic stimulation (rTMS) of the dorsolateral prefrontal cortex (DLPFC) has emerged as a promising antidepressant therapy. However, the heterogeneity of response underscores a pressing need for biomarkers of treatment outcome. We acquired resting state functional magnetic resonance imaging (rsfMRI) data in 47 MDD individuals prior to 5-8 weeks of rTMS treatment targeted using the F3 beam approach and in 29 healthy comparison subjects. The caudate, prefrontal cortex, and thalamus showed significantly lower blood oxygenation level-dependent (BOLD) signal power in MDD individuals at baseline. Critically, individuals who responded best to treatment were associated with lower pre-treatment BOLD power in these regions. Additionally, functional connectivity (FC) in the default mode and affective networks was associated with treatment response. We leveraged these findings to train support vector machines (SVMs) to predict individual treatment responses, based on learned patterns of baseline FC, BOLD signal power and clinical features. Treatment response (responder vs. nonresponder) was predicted with 85-95% accuracy. Reduction in symptoms was predicted to within a mean error of ±16% (r = .68, p < .001). These preliminary findings suggest that therapeutic outcome to DLPFC-rTMS could be predicted at a clinically meaningful level using only a small number of core neurobiological features of MDD, warranting prospective testing to ascertain generalizability. This provides a novel, transparent and physiologically plausible multivariate approach for classification of individual response to what has become the most commonly employed rTMS treatment worldwide. This study utilizes data from a larger clinical study (Australian New Zealand Clinical Trials Registry: Investigating Predictors of Response to Transcranial Magnetic Stimulation for the Treatment of Depression; ACTRN12610001071011; https://www.anzctr.org.au/Trial/Registration/TrialReview.aspx?id=336262).

Keywords: BOLD power; depression; functional connectivity; machine learning; magnetic resonance imaging; transcranial magnetic stimulation.

© 2019 Wiley Periodicals, Inc.

Figures

Figure 1

Methodology. MDD (n = 47) and HC (n = 29) were recruited and underwent MRI scanning. Following preprocessing, BOLD power was calculated at each voxel. Next, statistical comparison of MDD and HC data identified regions of significantly reduced BOLD power in the MDD group. Significant regions (p < 0.01) were binarized, generating a mask which was used to systematically calculate BOLD power for each individual and later examine predictors of rTMS treatment response. Maps of the default mode and affective network were generated by placing a seed in the posterior cingulate cortex and subgenual cingulate cortex respectively. An average group level connectivity map was then generated for each network. Individual BOLD power, AN and DMN connectivity constituted the neurobiological SVM features. MDD individuals underwent 5–8 weeks of initially daily (5 days per week, Monday to Friday) and later, titrated rTMS treatment, targeted to DLPFC using the F3 beam method as previously described (Bailey et al., 2018; Beam, Borckardt, Reeves, & George, 2009). After 3 weeks, responders continued to receive left‐sided treatment, whereas nonresponders were randomized to continue with left, right or bilateral treatment (full details in Supporting Information). Clinical severity was assessed using MADRS at baseline, Week 1, 3 and treatment endpoint. Initial clinical response at Week 1 was included as an optional SVM feature. SVM accuracy for prediction of treatment outcome was extensively interrogated using leave‐one‐out cross‐validation (CV), K‐fold CV, K‐fold CV with power and connectivity maps calculated de novo for each training sample permutation. SVM performance was additionally tested using SVM regression and the receiver operator characteristic curve. The manner in which the SVM separated R/NR was illustrated using the SVM hyperplane [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 2

Decreased power of resting‐state BOLD fluctuations in MDD. (a) Power was reduced in MDD compared to HC (p < .01; n = 41) in areas typically associated with reward circuitry, and was not increased in any region. Areas include bilateral mPFC, bilateral caudate and bilaterally a subregion of thalamus that is known to project to prefrontal cortex. (b) Three‐dimensional view of significant clusters showing a change in BOLD signal power (p < .01; family‐wise error corrected). (c) BOLD signal showing decreased power in MDD (blue) compared to HC (red; representative participants). (d) Power for each individual averaged within regions of significantly reduced power (bandwidth 0.01–0.1 Hz) and group average (bar line; *** p < .001). (e) BOLD frequency spectra from the identified regions for MDD (blue) and HC individuals (red). (f) Negative correlation between endpoint MADRS improvement and power in the identified regions (<0% response withheld [n = 8/33]: r = −.55; p = .004; all participants: r = −.42; p = .01). Individuals with a weak clinical rTMS response (<0%) are indicated in black circles with no fill [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 3

Affective network. (a) Within the AN, FC was significantly reduced in MDD compared to HC between the SGC seed (blue sphere) and a cluster comprising the medial PFC (p < .05; n = 43). (b) SGC‐mPFC connectivity is shown for each individual (asterisk represents p < .05). The MDD group showed greater heterogeneity compared to HC. (c) Spatial distribution of the affective network. (d) Connectivity within AN (Pearson's correlation coefficient, r) was associated with treatment outcome when individuals with lowest treatment response were withheld (<0% response withheld [n = 8/35]: r = −.40, p = .03; all participants: r = −.23, p = .17) [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 4

Treatment outcome prediction. (a) The cohort showed a bimodal distribution for clinical response to rTMS (% change in MADRS score). Distinct subpopulations display low and high treatment response outcomes, separated by a trough at 44.9% improvement. (b) SVM sensitivity, specificity and accuracy for binary classification of responders (R) and nonresponders (NR) was high, but varied slightly according to whether treatment response was defined according to (a) clinical outcome (>25% improvement) or (b) the trough separating subpopulations in the bimodal clinical response curve (45% improvement; n = 33). (c) Receiver operating characteristic (ROC) curves are displayed for binary classifier accuracy when R/NR were defined according to the clinical or bimodal split defined above. Both indicate performance in the “excellent” category (area under curve >0.9), although the bimodal split showed greater accuracy in identifying nonresponders. (d) SVM regression results indicated a good match between predicted and actual treatment response, although the SVM underestimated treatment outcome. (e) The relative contribution of each feature is illustrated by assessing accuracy when each is removed individually from the combined feature set. (f) The SVM hyperplane formed an elongated radial boundary encapsulating those individuals with the lowest power, DMN, and AN connectivity to classify R/NR [Color figure can be viewed at http://wileyonlinelibrary.com]