Global fMRI Harmonisation to Unlock Insights into Mechanisms, Mental Health and Neurostimulation

To create large, globally representative neuroimaging datasets for biomarker discovery we have to collaborate with and integrate data from institutes and data platforms across the world. But it’s not just the sample size that grows by increasing collaboration: working with distributed sites with often heterogeneous data practices and limited resources for harmonization can consume a lot of time and human effort, ultimately impeding research progress. The increasing international adoption of strong data privacy frameworks creates further uncertainty and inconsistency on what sites can share with each other. Together these challenges create a need for easy-to-use neuroinformatics tools that help create consistent data workflows at each site to enable an efficient, decentralized way of working together. In my talk I will focus on two projects I am involved in that each address one of these challenges: the Neurobagel project is an ecosystem for federated cohort discovery questions across distributed data sites such as “How many participants are there (across our network) who fit our cohort inclusion criteria and have been processed with freesurfer 7?”. It is built around the idea that data remain under the control of the collecting institution but are made discoverable by being harmonized through annotation with standardized terminology from existing, FAIR vocabularies. The Nipoppy project is a lightweight framework to standardize the curation and processing of an individual dataset. Its aim is to reduce the time and effort of adopting a new processing protocol or reprocessing data with an upgraded version of a pipeline consistently across many sites. Nipoppy builds on existing standards and tools for reproducible processing, and maintains a list of automatic extraction tools for data availability and imaging derived phenotypes for existing pipelines. I will present a practical case study of how these tools are integrated into existing collaborations, outline the plans for their ongoing development, and we will discuss both the successes and the social challenges that come with the adoption of these standards based, decentralized models of collaboration. 

Finding and validating biomarkers for mental health and behavior using neuroimaging requires combining many diverse cohorts in large-scale meta- and mega-analyses. In such studies, ensuring computational and analytic reproducibility becomes a core challenge. Five years ago, we released ENIGMA HALFpipe, a software built for collaborative projects creating such global benchmarks of biomarker discovery. HALFpipe is based on fMRIPrep for preprocessing and makes extensive use of other tools from the neuroimaging and open science ecosystem. At time of writing, more than ten resting state fMRI projects and three task-based fMRI projects are using HALFpipe within the consortium, with more than fifty thousand participants included across various psychiatric diagnoses and controls. Across these projects, we have encountered a number of important lessons that we will describe in this talk. Some examples of considerations include, for resting-state fMRI, the problem of analytic flexibility, and for task-based fMRI, how to model tasks on the individual level. For any task-based fMRI study, there are various ways in which researchers can operationalize a single psychological construct or cognitive domain into a concrete fMRI paradigm. We rely on an ontology-based protocol that standardizes design matrices and contrasts across sites. This ensures the interpretability of meta- and mega-analytic results. Another consideration involves the proportion of missing data we have encountered in meta- and mega-analyses. The brain coverage for different MRI scanners and acquisition protocols lead to missing parts of the brain in most multi-site datasets for at least some participants. This data is not missing at random, but is associated with head size and gender, meaning that statistics need to be carefully considered to avoid biased results. In this talk, I will discuss these and other challenges, as well as our roadmap for continuing to support large-scale collaborative investigations within ENIGMA and beyond. 

Major depressive disorder (MDD) is a leading contributor to the global burden of disease and disability. Despite extensive research, the neurobiological mechanisms underlying MDD remain poorly understood. Accumulating evidence suggests that MDD is characterised by alterations in functional connectivity, which may underlie the phenotypic manifestations of the disorder. However, findings from individual studies are often inconsistent, limiting the ability to draw definitive conclusions about patterns of connectivity alterations in MDD. These inconsistencies are likely due to the small sample sizes typical of neuroimaging studies, which result in limited statistical power, as well as substantial variability in the methods used to process fMRI data. In this talk, findings on functional alterations in MDD are presented, leveraging large-scale data from the ENIGMA MDD Consortium using standardised pipelines provided by the ENIGMA framework. The data includes resting-state fMRI scans from 25 cohorts worldwide analysed using a harmonised protocol (ENIGMA HALFpipe) to derive resting-state first-level features. These include connectivity metrics for specific seed regions (e.g., amygdala), large-scale networks (e.g., Default Mode Network), and measures of intrinsic neural activity (e.g., ReHo, fALFF). The resulting features were pooled using a mega-analysis framework to examine differences in resting-state functional connectivity between individuals with MDD and healthy controls. To account for variability across sites, linear mixed-effects models were employed with site included as a random effect. Results will be presented for the entire sample (MDD: N=2045, HC: N=2322, mean age = 39.7, SD = 17.14) as well as separately for adults (aged ≥25; MDD: N=1385, HC: N=1711) and youth (aged <25; MDD: N=660, HC: N=611). Across the whole sample, individuals with MDD showed lower connectivity compared to controls in specific regions, including the dorsolateral prefrontal cortex, posterior insula, thalamus, and fALFF. However, these differences were more pronounced when analysing adults and youth separately. Compared to controls, adults with MDD exhibited reduced connectivity between several seed regions (e.g., dorsomedial prefrontal cortex, precuneus, ventrolateral prefrontal cortex) and posterior brain regions (occipital cortex, lingual gyrus). In contrast, youth with MDD exhibited alterations predominantly in limbic and subcortical regions, including reduced connectivity between the amygdala and the insular cortex, the anterior insula and central operculum, the posterior insula and thalamus, and the thalamus and the precentral gyrus and anterior cingulate cortex. This comprehensive analysis highlights distinct connectivity alterations that vary by stage of development, providing new insights into the neurobiological underpinnings of MDD. The observed differences between adults and youth suggest that age-specific patterns of functional connectivity alterations may play a pivotal role in the neurobiology of MDD, highlighting the potential for tailoring neuromodulation treatments to target these distinct neural profiles more effectively. I will present these findings, and discuss the challenges and opportunities of large-scale multi-site rsfMRI harmonisation as well as potential implications of this work.  

The subgenual anterior cingulate cortex (sgACC) is central to the pathophysiology of major depressive disorder (MDD), with its functional connectivity (FC) to the left dorsolateral prefrontal cortex (DLPFC) influencing outcomes of transcranial magnetic stimulation (TMS). However, inconsistent findings in sgACC-FC research often result from small sample sizes and site-specific variations, underscoring the need for harmonised, large-scale analyses. Leveraging the Depression Imaging REsearch ConsorTium (DIRECT) Phase II—a globally harmonised fMRI dataset with 1660 MDD patients and 1341 healthy controls—we systematically delineated sgACC-FC abnormalities and their implications for personalised TMS targeting. Enhanced sgACC-DLPFC FC in MDD patients relative to healthy controls shifted the sgACC anti-correlation peak within the DLPFC, altering the optimal TMS target location. Using these insights, we developed an MDD big data-guided TMS targeting algorithm that integrates group-level FC maps with individual-level brain activity for precise, patient-specific TMS target identification. This symposium talk will highlight the advantages of harmonised fMRI datasets in overcoming site variability, enhancing statistical power, and driving clinical applications. We demonstrate how harmonised, cross-cohort analyses improve our understanding of MDD-related FC disruptions and provide a foundation for actionable biomarkers. Furthermore, our algorithm’s validation across independent clinical samples shows its potential to improve TMS outcomes through personalised targeting.