Systematic fMRI signal differences across cohorts alter lifespan connectome trajectories

Abstract

Large-scale lifespan neuroimaging studies increasingly integrate data across distinct cohorts to characterize trajectories of brain development and aging. However, systematic differences in acquisition protocols and hardware across cohorts can alter signal characteristics in ways that bias downstream analyses. Here, we examine three cohorts from the Human Connectome Project (HCP), spanning development (HCP-D), young adulthood (HCP-YA) and aging (HCP-A), to illustrate this issue and evaluate existing strategies to mitigate it. HCP has set standards for open, deeply phenotyped, high-resolution human neuroimaging, which are frequently used as high-quality reference datasets in tool validation, replication studies, and cross-cohort meta-analyses. However, neuroimaging acquisitions have differed across HCP cohorts because of changes in scanner hardware and acquisition sequences across study phases. Because of HCPs widespread usage, even modest protocol differences between cohorts-and their downstream effects-can have outsized impacts on the field of neuroscience research. Our analysis reveals that the HCP-YA cohort exhibits systematically weaker temporal signal-to-noise ratio (tSNR) relative to HCP-D/A. These signal quality discrepancies propagate to downstream analyses, leading to differences in overall resting-state functional correlations and whole-brain and node-level measures of resting-state network organization (e.g., system segregation, modularity, participation coefficient). Consistent with protocol-driven signal differences, resting-state network measures derived from HCP-YA depart from expected lifespan trajectories, as confirmed by examination of two other lifespan datasets. Harmonization approaches accounting for protocol and scanner-model differences substantially lessen these artifactual differences in brain network measures. Our findings underscore that signal differences do not merely introduce noise, but can qualitatively alter estimated lifespan trajectories of functional network organization, including partially inverting expected lifespan patterns. Without appropriate harmonization, analyses that combine HCP cohorts can therefore result in biologically misleading inferences about brain development and aging. We demonstrate how small acquisition differences bias resting-state-derived network metrics, and how these effects can be mitigated. This work advances best practices for valid inference in multi-cohort lifespan neuroscience research.