Harnessing mega-analysis in the era of “big data” neuroimaging

Associations between the brain and psychiatric symptoms are likely to be small [1]. Detecting these associations reliably requires large sample sizes, beyond the reach of single research sites. Traditionally, meta-analyses of published imaging studies have been used to aggregate data. However, these approaches are vulnerable to publication bias and synthesize heterogeneous group-level differences that achieve significance in small, underpowered samples, which are susceptible to both type-I and type-II errors [1, 2].

A potentially more robust approach to data aggregation is neuroimaging mega-analysis, initially popularized by the Enhancing Neuro Imaging Genetics through Meta Analysis (ENIGMA) consortium [3]. This method involves the uniform processing and integration of individual participant-level data across multiple datasets, containing thousands of participants. Initial work from ENIGMA reported smaller subcortical volumes and cortical surface area in youth with ADHD compared with controls [3]. We have applied such mega-analytic methods to data from large publicly available datasets, resulting in a combined dataset of ~9000 youth with or without Attention-Deficit/Hyperactivity Disorder (ADHD). Our findings support prevailing neurobiological models, revealing greater functional connectivity between the default mode network and task-positive networks [1]; greater functional connectivity within subcortico-cortical circuits centered on the caudate [2]; and less fractional anisotropy within long association white matter tracts [4], in youth with ADHD compared to unaffected controls. Similar associations were found with ADHD-traits in the broader population [1, 2, 4].

This mega-analytic approach offers important advantages over traditional meta-analyses (Fig. 1). First, our investigations reported small yet statistically significant associations, with maximum effect sizes of d = 0.21 for neuroanatomy and d = 0.17 for functional connectivity [1, 3]. These effects likely would be undetectable via traditional meta-analyses of published findings, which are limited to analyzing clusters meeting statistical significance in heterogenous underpowered studies [1]. Furthermore, owing to the large participant numbers in mega-analyses and the use of participant-level data, we were able to match participants with and without ADHD on the degree of in-scanner motion, removing a major participant-level confound in neuroimaging [1, 2]. Similarly, by matching participants according to comorbid internalizing and externalizing behaviors, we have demonstrated that ADHD-associated differences in functional connectivity are not driven by comorbid emotional and behavioral problems [1]. Such matching and consideration of individual characteristics is not possible in meta-analyses of published group-level results.

Fig. 1. Comparison of neuroimaging meta-analysis and mega-analysis approaches in ADHD research.

a A graphic of neuroimaging meta-analysis. b A graphic of neuroimaging mega-analysis along with some of its advantages. The green and blue represent those with and without ADHD respectively from 1 to n studies. Key: d = effect sizes from studies 1 through n.

Can such small effects be translated clinically? One promising approach involves aggregating brain-behavior associations into cumulative “polyneuro scores”, analogous to polygenic risk scores. Initial work indicates that these scores can be prognostic, linked to long-term outcomes, such as distinguishing participants with persistent ADHD from unaffected controls [5]. Mirroring the increases in variance explained by polygenic scores based on large-scale, consortia-led genome-wide association studies, the adoption of a mega-analytic approach in the calculation of polyneuro scores is likely to improve their predictive power. Furthermore, the large sample sizes and potential for rigorous cross-validation in mega-analyses may support the identification of robust, reproducible biological subtypes. For instance, recent mega-analytic work identified subtypes of major depression based on functional connectivity patterns, which predicted responses to antidepressants [6]. A similar strategy could be beneficial for ADHD.

Author contributions

LJN and PS wrote the paper.

Funding

Funded by the NIMH Intramural Research Program and the National Human Genome Research Institute (grant ZIAHG200378 to Dr. Shaw).

Competing interests

The authors declare no competing interests.