Gray matter correlates of MDD: a cross-cohort investigation of replicability and generalizability

499 

Abstract Submission 

Janik Goltermann¹, Udo Dannlowski², Klaus Berger³, Tilo Kircher⁴, Tim Hahn²

¹University of Münster, Münster, NRW, Germany, ²Institute for Translational Psychiatry, Münster, North Rhine Westphalia, ³Institute of Epidemiology and Social Medicine, University of Münster, Münster, NRW, Germany, ⁴University of Marburg, Marburg, Germany

Janik Goltermann  
University of Münster
Münster, NRW, Germany

Udo Dannlowski  
Institute for Translational Psychiatry
Münster, North Rhine Westphalia

Klaus Berger  
Institute of Epidemiology and Social Medicine, University of Münster
Münster, NRW, Germany

Tilo Kircher  
University of Marburg
Marburg, Germany

Tim Hahn  
Institute for Translational Psychiatry
Münster, North Rhine Westphalia

Major depressive disorder (MDD) is one of the leading causes of disability worldwide (1) and is still insufficiently treated, with approximately one third of patients being treatment-resistent (2). Despite the societal relevance of depression and a plethora of research over the past decades, the neurobiological underpinnings of the disorder are still poorly understood. Previous neuroimaging studies have yielded highly heterogeneous results, even across large consortia (3–5) and effect sizes appear to be subtle at most (6). In addition, the validity of brain-behavior findings in general have been questioned due to reports of underpowered study samples, overestimated effect sizes and overall low replicability (7). These findings make it highly relevant to systematically investigate the replicability of the neural correlates of psychiatric disorders, such as depression and assess their generalizability to independent cohorts.

Three independent cohorts totaling N=4021 adult participants were analyzed, including lifetime MDD patients (n=1764) and health control (HC) individuals (n=2257). Cohorts were from the MACS (n=1799), the MNC (n=1198) and the BiDirect study (n=1024). Lifetime MDD diagnosis was determined using structured interviews, conducted by clinically trained raters. Brain-wide association between MDD and gray matter voxel-based morphometry was tested using general linear models, controlling for age, sex and total intracranial volume. Cross-cohort replicability was assessed by inspecting the congruence of significant voxels between the three cohorts, at a liberal uncorrected threshold of p<.001 and using a stringent FWE-corrected threshold of pFWE<.05. Cross-cohort generalizability was investigated using a cross-validation framework, iterating through the three cohorts as independent test sets.

Using an uncorrected threshold of p<.001 an overlap in significance congruently between all three cohorts was found in clusters distributed mainly across the thalamus, insula, as well as the lingual, fusiform and parahippocampal gyri (a total of k=787 voxels; Figure 1). When correcting for multiple comparisons at pFWE<.05 no voxels showed significant effects congruently across all three cohorts. However, all pairwise combinations of cohorts showed an overlap in congruent significance even at this level in some areas, located mainly within the thalamus, insula, orbitofrontal cortex, and the fusiform and lingual gyri. All observed effect sizes were small, with maximum explained variance of the diagnosis effect up to partial R²=.01. Inspecting the threshold-free cohort-wise generalizability of the diagnosis effect using cross-validation yielded largely generalizable effects across cohorts (Figure 2). Largest and most robust generalization was achieved within bilateral insulae. Effects within the parahippocampal-lingual compound, the thalamus, the cerebellum, and widespread prefrontal clusters generalized between some cohort combinations but not others.

Gray matter correlates of depression are in part replicable and even generalizable between well-powered independent cohorts, despite the small magnitude of effects. Evidence implies a network containing the insula, thalamus and a lingual-parahippocampal complex. Most robust effects were found for the insula, while little evidence was found for a replicable or generalizable effect located within the hippocampus, opposed to a variety of previous findings. The most important limitation is the relatively low demographic diversity across cohorts (German, caucasian and comparably high socioeconomic status). Nevertheless, our findings demonstrate the need of replication efforts and the potential utility of cross-validation for univariate analyses in order to ensure the generalizability of findings, thus counteracting current concerns of low replicability.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Methods Development

Univariate Modeling ²

Cortical Anatomy and Brain Mapping

Subcortical Structures

Affective Disorders

Data analysis

Psychiatric Disorders

Statistical Methods

STRUCTURAL MRI

Structures

Sub-Cortical

Univariate