Normative Modeling in Neuroimaging

• Normative modeling in neuroimaging is a statistical framework that computes individual deviation scores by comparing brain measures against a healthy reference distribution adjusted for covariates.
• It integrates diverse methodologies from linear regression to deep learning to capture complex, non-linear and high-dimensional imaging data effectively.
• This approach supports individualized risk assessment and anomaly detection, facilitating precision diagnostics and disease stratification in clinical research.

Normative modeling in neuroimaging is a statistical and computational framework designed to quantitatively characterize the expected variability of brain measures in a reference (typically healthy) population, enabling subject-specific quantification of deviations from these population-level distributions. Unlike traditional case-control approaches, which detect group-average differences, normative modeling focuses on individualized “growth chart–like” reference mapping, providing a flexible, covariate-adjusted benchmark for the identification of structural or functional abnormalities across heterogeneous populations.

1. Core Concepts and Rationale

Normative modeling learns the distribution of brain measures—such as cortical thickness, gray matter volume, network connectivity, or functional activation—conditional on covariates (e.g., age, sex, scanner site) from large-scale healthy samples. The key output is an individual deviation score (e.g., z-score or centile), quantifying how a new individual’s metric(s) compare to what would be expected given their relevant covariates. This approach has several critical advantages over case–control designs:

• Subject-level inference: Each individual is evaluated against the normative reference distribution, enabling the detection of idiosyncratic abnormalities even within highly heterogeneous diseases.
• Covariate adjustment: Statistical adjustment for known sources of variation (e.g., age, sex, site) is built into the model, yielding deviation estimates that are robust to demographic or technical confounding (Alyas et al., 8 Sep 2025).
• No need for large, matched controls: Once a normative model is trained on a large healthy reference set, new clinical samples from a distinct site or population can be assessed after calibration, obviating the need for retraining or recruiting new controls for every comparison (Little et al., 3 Jun 2024, Alyas et al., 8 Sep 2025).

Mathematically, the deviation score typically takes the form

$$z_i = \frac{y_i - \hat{\mu}(x_i)}{\hat{\sigma}(x_i)}$$

where $y_i$ is the individual's measurement, $\hat{\mu}(x_i)$ and $\hat{\sigma}(x_i)$ are the predicted mean and standard deviation from the normative model at covariate values $x_i$.

2. Statistical and Machine Learning Methodologies

Normative modeling approaches employ a hierarchy of statistical and machine learning models with escalating complexity and flexibility, tailored to the properties of neuroimaging data:

Methodology	Strengths	Limitations
Linear Regression / Polynomial	Simple, interpretable, covariate-adjusted	May not capture nonlinear, heteroscedastic effects
Gaussian Process Regression (GPR)	Flexible, robust to nonlinearity, yields predictive uncertainties	Computationally challenging for high-dimensional data (Kia et al., 2018)
Bayesian/Hierarchical Regression	Regularizes site/subject effects, enables recalibration	Computationally intensive, sensitive to hyperprior specification (Kia et al., 2020)
Generalized Additive Models (GAMLSS)	Captures non-Gaussianity (skew, kurtosis), nonlinear covariate effects	May be limited in very high-dimensional imaging spaces (Little et al., 3 Jun 2024, Hu et al., 3 Jun 2025)
Deep Neural Networks/VAEs/Transformer models	Models complex, high-dimensional dependencies; scalable to whole-brain	Requires large datasets, more challenging interpretability (Costa et al., 2022, Kia et al., 2018, Aguila et al., 2023)

Multimodal and tensor approaches—such as multi-task GPR tensor regression (Kia et al., 2018), deep variational autoencoders (Aguila et al., 2023, Kumar et al., 2021), and transformer-based generative models (Costa et al., 2022)—have been adopted to address the distinct challenges posed by the multi-way, structured, and high-dimensional nature of neuroimaging data. Key innovations include:

• Kronecker and tensor algebra to model and efficiently compute structured covariance (across subjects and space) (Kia et al., 2018, Kia et al., 2018)
• Low-rank and manifold learning (e.g., PCA, latent variable models) to reduce output dimensionality
• Conditional and adversarial VAEs for disentangling covariate and pathological sources of variance, improving out-of-distribution generalization (Wang et al., 2022)
• Product/Mixture-of-Experts strategies to fuse information from multiple modalities or sources (Kumar et al., 2023, Aguila et al., 2023)
• Scalable deep generative modeling (transformers, diffusion, graph VAE) to capture subtle, spatially distributed or network-level changes (Costa et al., 2022, Zhang et al., 7 Mar 2024, Shen et al., 14 Oct 2024, Ijishakin et al., 19 Jul 2024)

3. Handling High-Dimensional, Multi-Modal, and Multi-Site Data

Normative modeling in neuroimaging must address unique challenges of data dimensionality, heterogeneity, and technical variation:

• High dimensionality: Imaging data often have tens of thousands of features (voxels, vertices, network edges). Techniques such as low-rank approximations, tensor decompositions, and latent-space modeling reduce the number of parameters and ensure computational tractability without discarding critical spatial or network structure (Kia et al., 2018, Kia et al., 2018, Aguila et al., 2023).
• Multimodal integration: Across MRI, PET, EEG/MEG, and neuroimaging-based ATN biomarker sets, multimodal VAEs and Mixture-of-Product-of-Experts strategies allow joint modeling of disparate features, capturing both complementary and redundant sources of variation (Kumar et al., 2023, Kumar et al., 4 Apr 2024).
• Site/scanner differences: Hierarchical Bayesian models implement partial pooling, sharing statistical strength across sites while regularizing site-specific model parameters (Kia et al., 2020, Alyas et al., 8 Sep 2025). Empirical evidence indicates that calibration with ∼30 local controls is sufficient for robust out-of-sample deviation estimates.
• Non-Gaussianity and outlier modeling: GAMLSS and skew-normal distributions allow the estimation of normative ranges in presence of skew and heavy-tailed outlier distributions typical of real-world neuroimaging datasets (Little et al., 3 Jun 2024, Palma et al., 8 Jul 2024).

A selection of open access toolkits embedding these methodologies includes BrainChart, Brain MoNoCle, PCN Toolkit, and CentileBrain, each leveraging different statistical back-ends and providing pre-trained models for direct application to new data (Little et al., 3 Jun 2024, Alyas et al., 8 Sep 2025).

4. Applications: Deviation Quantification and Novelty Detection

The output of a normative model is an individualized quantification of abnormality, typically via region-wise z-scores, centiles, or global deviation indices. These provide:

• Region-wise abnormality mapping: Z-maps or centile maps highlight specific brain regions or networks where individuals deviate from the norm, supporting anatomical localization and hypothesis-driven analysis (Palma et al., 8 Jul 2024, Kumar et al., 2021).
• Disease severity indices: For multimodal models (e.g., ATN biomarkers), the spatial extent and magnitude of deviations can be aggregated into summary indices (e.g., Disease Severity Index, DSI) that correlate with clinical scales and risk of disease progression (Kumar et al., 4 Apr 2024).
• Individualized risk assessment: Extreme deviation indices, such as the mean of the absolute z-scores above a high quantile, are associated with neuropsychological outcomes or risk of disease (Palma et al., 8 Jul 2024, Ijishakin et al., 19 Jul 2024).
• Novelty/anomaly detection: In unsupervised settings, the sensitivity and specificity of normative models to rare or subtle abnormalities (psychiatric, neurodegenerative, epileptic) can surpass traditional classification approaches (Kia et al., 2018, Costa et al., 2022, Kia et al., 2018).
• Cross-modal and network divergence: Recent frameworks enable subject-level quantification of anomalous whole-brain EEG network topology (Hu et al., 3 Jun 2025), connectomic developmental deviations (Shen et al., 14 Oct 2024), and subject-specific outlier connectivity patterns.

5. Model Calibration, Site Effects, and Workflow Considerations

Practical implementation of normative modeling requires consideration of key factors for validity and reproducibility:

• Calibration: For accurate deviation quantification, each new cohort or site should be calibrated using a modest-sized, demographically-matched (ideally site-specific) subsample from the local data (Alyas et al., 8 Sep 2025, Kia et al., 2020). Site-mismatched controls or protocol incompatibilities can introduce systematic bias and large effect size discrepancies.
• Choice of statistical platform: While the relative patterns of abnormality are highly consistent across tools (agreement in Cohen’s d), absolute z-scores may have platform-dependent scaling. For rigorous studies, outputs should be cross-validated across multiple modeling frameworks (Alyas et al., 8 Sep 2025).
• Quality control and decision points: Transparent documentation of workflow parameters (e.g., segmentation, referencing, metric selection) is critical in EEG/MEG and intracranial electrophysiology analyses (Woodhouse et al., 6 Feb 2025, Hu et al., 3 Jun 2025). Modular guides detail standard practices for metric selection, segment definition, outlier rejection, and database handling.

6. Future Directions and Methodological Innovations

Current and emerging trends in normative modeling include:

• Advanced generative architectures: Increased adoption of transformer-based autoregressive models, conditional diffusion models (notably in the non-Euclidean/surface domain), and deep graph VAEs for structural and functional connectomes (Costa et al., 2022, Zhang et al., 7 Mar 2024, Shen et al., 14 Oct 2024, Ijishakin et al., 19 Jul 2024).
• Non-Gaussian and non-linear expansions: Incorporation of flexible distributional models (sinh–arcsinh, skew-normal) for more faithful characterization of real imaging measures, especially in aging and disease (Little et al., 3 Jun 2024, Palma et al., 8 Jul 2024).
• Data integration and personalization: Models integrating genetic, behavioral, and longitudinal clinical data are under exploration, facilitating individualized prediction and intervention (Kumar et al., 2023, Kumar et al., 4 Apr 2024).
• Open access and clinical translation: Pre-trained, user-friendly normative modeling platforms are becoming widely available with web-based interfaces, allowing routine application in both research and clinical practice (Little et al., 3 Jun 2024, Alyas et al., 8 Sep 2025).

A plausible implication is that as cohort sizes and data modalities continue to grow, normative modeling frameworks that articulate individual risk profiles and spatiotemporal deviation maps will underpin precision diagnostics and monitoring in both neurology and psychiatry.

7. Limitations and Methodological Pitfalls

Several critical caveats and challenges remain:

• Sample representativeness and generalizability: If the reference sample is not representative of the target population (in terms of age, sex, ancestry, etc.), model deviation scores may be misleading. Proper covariate adjustment and calibration are necessary.
• Residual confounding: Although sophisticated models (hierarchical Bayes, GAMLSS) mitigate site and demographic effects, technical variability or unmeasured confounders may still distort deviation estimates (Kia et al., 2020, Alyas et al., 8 Sep 2025).
• Interpretability and causality: Deep learning models provide flexible mapping but can obscure causal or mechanistic interpretation. Ensuring inter-model agreement and explaining individual deviations remain significant challenges (Eitel et al., 2023).
• Bias–variance trade-off: Overly complex models may overfit the training data. Regularization strategies (e.g., low-rank approximations, variance shrinkage, informative priors) are essential.
• Calibration cohort size: Very small calibration cohorts (<10) may yield unstable site-effect estimates, and mismatched calibration can introduce systematic errors (Alyas et al., 8 Sep 2025).

Robust implementation thus requires careful attention to data provenance, calibration, choice of modeling framework, and clear documentation of all analytical decisions.

---

Normative modeling in neuroimaging has evolved into a foundational approach for subject-level anomaly detection, disease stratification, and individualized biomarker discovery. Its methodological diversity—spanning classic regression, advanced Bayesian statistics, and deep generative modeling—continues to expand as novel data types, computational resources, and clinical demands arise. By providing covariate-adjusted individual deviation scores and enabling fine-scale localization across high-dimensional multi-modal data, normative modeling serves as a critical scaffold for the next generation of neuroimaging research and clinical translation (Alyas et al., 8 Sep 2025, Little et al., 3 Jun 2024, Kia et al., 2020, Kia et al., 2018).