Brain turbulence metrics shift with age, sex, and scan motion in a UK Biobank cohort
A large resting-state fMRI analysis has put a hard number on a problem that has hovered over turbulence dynamics research for years: the signal is not purely biological. In 11,184 participants from the UK Biobank, age, sex, and head motion during scanning were each statistically significantly associated with all turbulence dynamics features examined across 11 spatial scales, even though the effect sizes were generally small, with partial eta squared below 0.01. Those weak effect sizes should not be mistaken for irrelevance. In imaging science, a small systematic bias across thousands of brains can distort the interpretation of a supposedly disease-related readout.
The work focused on the turbulence dynamics framework, which has been used to characterize whole-brain network synchronicity and rigidity and has already been linked to neuropsychiatric disorders and antidepressant response prediction. The challenge is familiar to anyone working with large neuroimaging cohorts: demographic factors and acquisition artifacts can seep into the feature space and make a disease signal look stronger, weaker, or simply different from what it really is. By testing associations across spatial scales from λ 0.01 to 0.31, the analysis makes clear that the readouts are sensitive to who is in the scanner and how they move inside it.
Older age tracked with increased network rigidity. That finding fits the broader clinical intuition that the brain’s dynamic repertoire tends to narrow with aging, with less flexible coordination across distributed circuits. In practical terms, a model built to infer neuropsychiatric status from resting-state dynamics cannot assume that the same value means the same thing in a 28-year-old and a 72-year-old. Sex differences and motion effects were also present across the full set of turbulence features, reinforcing the need to treat these covariates as part of the signal-processing problem rather than as background noise.
Normative modeling corrected the signal without improving depression prediction
Investigators then asked whether normative modeling could correct for demographic variability and improve prediction of recent depressive symptoms measured with the RDS-4 questionnaire. It did not increase overall predictive power. The depressive symptom phenotype appears to have been weak, and that likely capped any measurable gain from covariate correction. A flat outcome on prediction does not erase the value of the correction step, because the key question was not only whether the model could classify symptoms better, but whether it could recover a more disease-relevant version of the turbulence signal.
Here the result was more interesting than the headline prediction metric. After correction, the most predictive spatial scale for Amplitude Turbulence shifted to an intermediate scale, λ = 0.16. That relocation matters because it aligns the analysis with prior turbulence literature on health and disease, suggesting that normative adjustment stripped away some of the demographic and motion-related noise that can obscure the biologically meaningful scale. In other words, the model did not become more powerful in a raw forecasting sense, but it became more literate in the language of the underlying neurobiology.
The value of that shift is easy to understand at the bedside and in the imaging lab alike. If a biomarker is sensitive to age and motion, the most predictive scale may simply be the scale most exposed to confounding. A normative model can pull the estimate back toward a scale that better reflects the pathophysiology researchers are trying to study, especially in disorders where distributed network dynamics are central to symptom expression.
Why small confounding effects still matter in whole-brain imaging
The study’s central warning is straightforward: turbulence dynamics features are not demographically neutral. Age, sex, and head motion during scanning all left statistically significant fingerprints on the measures, even though each effect was small. In a dataset this large, small effects become operationally important because they can accumulate across analyses, be amplified in subgroup comparisons, or distort replication attempts across centers and scanners.
Head motion deserves special attention because it is a perennial source of trouble in resting-state fMRI. Motion can mimic or obscure coherence patterns, and any feature that aims to summarize brain-wide synchrony will be vulnerable if scan quality is uneven. Correcting for motion alongside age and sex is therefore not just a statistical courtesy; it is part of the basic hygiene needed before claims about disease specificity can be trusted.
The authors frame the main payoff of normative modeling not as a prediction boost but as a way to recover a more literature-consistent signal. That is a restrained and defensible conclusion. When a biomarker is still being defined, fidelity to the underlying biology may matter more than a marginal gain in classification metrics, especially when the phenotype being predicted is noisy, transient, or weakly expressed.
The clinical and research implications of a more disease-specific signal
This work lands in a domain where the translational promise is high but the analytic traps are numerous. Resting-state fMRI has long been attractive because it offers a noninvasive window into systems-level brain organization, and turbulence dynamics adds a mechanistic framing around rigidity and synchrony rather than just correlation matrices. Yet any measure that aims to track antidepressant response or neuropsychiatric burden must survive the reality that brain imaging is also a record of age, sex, and movement.
The proposed normative model offers a cleaner way to interpret turbulence dynamics in future studies. It does not solve the problem of weak symptom phenotypes, and it does not create a new predictive leap where none existed. What it does do is sharpen the map by reducing the influence of factors that are not the disease itself. For a field trying to separate state effects from subject-level variability, that is a meaningful step in measurement discipline.
The remaining challenge is methodological rather than conceptual. Larger studies will still need to determine whether corrected turbulence features replicate across different psychiatric cohorts, whether the same spatial scale emerges in clinically defined depression, and how stable these findings are across scanners, preprocessing pipelines, and symptom measures. The current analysis shows that the brain’s turbulence readouts are sensitive to who is being measured and how the data are acquired. Any future biomarker claim will have to clear that bar before it can be treated as clinically trustworthy.