Predicting Antidepressant Treatment Response From Cortical Structure on MRI: A Mega-Analysis From the ENIGMA-MDD Working Group

Abstract

Accurately predicting individual antidepressant treatment response could expedite the lengthy trial-and-error process of finding an effective treatment for major depressive disorder (MDD). We tested and compared machine learning-based methods that predict individual-level pharmacotherapeutic treatment response using cortical morphometry from multisite longitudinal cohorts. We conducted an international analysis of pooled data from six sites of the ENIGMA-MDD consortium (n = 262 MDD patients; age = 36.5 ± 15.3 years; 154 (59%) female; mean response rate = 57%). Treatment response was defined as a ≥ 50% reduction in symptom severity score after 4-12 weeks post-initiation of antidepressant treatment. Structural MRI was acquired before, or < 14 days after, treatment initiation. The cortex was parcellated using FreeSurfer, from which cortical thickness and surface area were measured. We tested several machine learning pipeline configurations, which varied in (i) the way we presented the cortical data (i.e., average values per region of interest, as a vector containing voxel-wise cortical thickness and surface area measures, and as cortical thickness and surface area projections), (ii) whether we included clinical data, and the (iii) machine learning model (i.e., gradient boosting, support vector machine, and neural network classifiers) and (iv) cross-validation methods (i.e., k-fold and leave-one-site-out) we used. First, we tested if the overall predictive performance of the pipelines was better than chance, with a corrected 10-fold cross-validation permutation test. Second, we compared if some machine learning pipeline configurations outperformed others. In an exploratory analysis, we repeated our first analysis in three subpopulations, namely patients (i) from a single site, (ii) with comparable response rates, and (iii) showing the least (first quartile) and the most (fourth quartile) treatment response, which we call the extreme (non-)responders subpopulation. Finally, we explored the effect of including subcortical volumetric data on model performance. Overall, performance predicting antidepressant treatment response was not significantly better than chance (balanced accuracy = 50.5%; p = 0.66) and did not vary with alternative pipeline configurations. Exploratory analyses revealed that performance across models was only significantly better than chance in the extreme (non-)responders subpopulation (balanced accuracy = 63.9%, p = 0.001). Including subcortical data did not alter the observed model performance. Cortical structural MRI alone could not reliably predict individual pharmacotherapeutic treatment response in MDD. None of the used machine learning pipeline configurations outperformed the others. In exploratory analyses, we found that predicting response in the extreme (non-)responders subpopulation was feasible on both cortical data alone and combined with subcortical data, which suggests that specific MDD subpopulations may exhibit response-related patterns in structural data. Future work may use multimodal data to predict treatment response in MDD.

Keywords: ENIGMA; Radiomics; antidepressant treatment response; machine learning; magnetic resonance imaging; major depressive disorder; mega‐analysis.

FIGURE 1

Processing and analysis pipeline. The complete analysis pipeline is presented from left to right, starting with data acquisition and preprocessing steps. Four steps, part of the machine learning pipeline, follow these. We tested various machine learning pipeline configurations for the four steps presented in our secondary analyses. Finally, the model is trained and tested, initially on the full population and subsequently in exploratory analyses on three subpopulations. Finally, we test if adding subcortical data improves predictive performance for our secondary exploratory analysis. ENIGMA MDD, Enhancing Neuroimaging Genetics through Meta‐Analysis Major Depressive Disorder Working Group; GBC, Gradient boosting classifier; imp., imputation; sel, feature selection; SVC, support vector classifier; reg., regressing out confounders; ROI, Region of interest; T1w MRI, T1‐weighted Magnetic Resonance Imaging.

FIGURE 2

CONSORT flow diagram of patient inclusion.

FIGURE 3

Region of interest relevance for the exploratory analyses. The figure shows the normalized coefficients used by a machine learning model to predict treatment response in the extreme (non‐)responders subpopulation, overlaid on a standard structural brain (cortical data representation: ROI average; model: gradient boosting classifier [GBC]; only cortical data). The sign indicates the direction of the relationship between a positive treatment response and either surface area (left two panels) or cortical thickness (right two panels). Red indicates a positive direction, whilst blue indicates a negative direction. The magnitude (visualized as saturation) of the coefficients indicates the strength of the relationship.