Large-scale population neuroimaging reveals latent subgroup structure in functional brain organisation

Researchers analyzed brain scans from nearly 20,000 people to find hidden groups of people who share similar brain patterns. They discovered that these brain groups are not random. They are actually linked to how people think, their lifestyle, their physical health, and even their DNA.

In neuroimaging, scientists use functional MRI (fMRI) to record how brain regions communicate while a person is at rest. Traditionally, researchers have focused on the "average brain." They create consensus models that describe how a typical human brain functions. While this helps us understand basic mechanics, it ignores the massive, structured variation between individuals. This creates a gap. We have models for the "average" person and data for "specific" individuals. However, we lack a way to understand the organized subgroups that sit between them.

Note for clinicians: This study identifies patterns in large populations. It is not a validated tool for diagnosing individual patients in a clinical setting.

Bridging the population-individual gap

Most existing neuroimaging approaches struggle to move from population averages to individual specificity. Conventional models often rely on small, homogeneous cohorts. These usually involve fewer than 50 people. This makes it difficult to see the subtle, structured heterogeneity in a diverse population. Even when researchers use large datasets, they often focus on "functional connectivity." This measures the temporal correlations (how the timing of activity in one region relates to another) between predefined brain parcels.

While connectivity is a powerful tool, the authors argue it misses a critical dimension: the spatial organization of functional modes. Think of functional modes as the "geographic" layout of brain activity. Knowing the layout means knowing exactly where communication hubs are located, not just when regions talk to each other. Previous methods failed to bridge these two worlds. Researchers were unable to identify how specific, latent subgroups might differ in their actual brain architecture.

From raw voxels to functional fingerprints

To solve this, the authors developed a scalable framework for unsupervised subgroup discovery. This means the system identifies patterns without being told what to look for. The pipeline, illustrated in, uses several stages of dimensionality reduction to handle massive data from 19,993 UK Biobank participants.

1. Estimating Individual Topographies: The process begins with stochastic Probabilistic Functional Modes (sPROFUMO). This is a hierarchical Bayesian matrix factorization model. You can think of matrix factorization like breaking a complex recipe down into its individual ingredients. It estimates individualized spatial maps for each participant. This creates a custom "map" of their resting-state networks.
2. Extracting Fingerprints: These maps are processed through two complementary methods: bigFLICA and subjectICA. bigFLICA extracts shared patterns of spatial variation across multiple functional modes. Meanwhile, subjectICA focuses on finding subject-independent components. Together, these produce 1,000 "fingerprint features." These are mathematical descriptions of an individual's unique brain organization.
3. Discovering Subgroups: Finally, the researchers apply 1D Gaussian Mixture Modelling (GMM) to each fingerprint feature independently. This approach looks for "non-Gaussianities" (departures from a standard bell curve). If a feature shows a long tail (skewness) or heavy ends (kurtosis), the GMM identifies those outliers as distinct, latent subgroups .

Thousands of links to human traits

The scale of the discovery is significant. The authors report approximately 5,700 significant differences between these discovered subgroups. These differences span a wide array of non-imaging phenotypes (observable traits). These include categories like cognition, mental health, lifestyle, and physical health .

Crucially, the study finds that these subgroups correspond to real-world biological variation. For example, the authors report that subgrouping provides stronger brain-phenotype associations in five out of six categories. This includes cognitive and mental health categories [Figure 3b]. This is better than using continuous fingerprint features alone. This suggests that categorizing people into these latent subgroups actually reduces noise. It also captures the most relevant biological signals.

Furthermore, the spatial organization of these subgroups follows a distinct "block structure" [Figure 4a]. The differences found in sensory-motor networks (like vision and movement) are highly similar to one another. They are distinct from the differences found in higher-order cognitive networks (like language and attention). The study also demonstrates a biological link to genetics. The spatial patterns of subgroup variability correspond to regional patterns of genetic variation across the brain .

Limits of the unsupervised approach

There are important boundaries to what this framework can claim. First, the study identifies strong associations between brain subgroups and traits. However, the authors explicitly state they have not identified causal mechanisms. The framework tells us that certain brain patterns cluster with certain behaviors. It does not tell us why those patterns emerge.

Second, the entire discovery process is unsupervised. This prevents researcher bias. However, it also means the system relies heavily on the presence of non-Gaussianities in the data. If a biological trait manifests as a perfectly smooth, Gaussian distribution, this framework might fail to identify it as a distinct subgroup. Finally, while the results are highly reproducible across data splits, the framework is optimized for discovering "latent" structure. It is not yet a tool for predicting a specific individual's health outcome.

A new foundation for stratified medicine

This work is a foundational proof-of-concept for population neuroscience. It is not yet a diagnostic tool for individual patients. However, it provides a vital bridge for the future of precision medicine.

By demonstrating that large-scale fMRI data contains reproducible subgroups, the authors have provided a roadmap. Future clinical research can use these latent subgroups to categorize patients into more biologically accurate cohorts. This could eventually allow for more targeted interventions. These would be tailored to the specific functional architecture of a patient's brain.

Figures from the paper