🧠 neuroscience

Protocol Update: The Normative Modelling Paradigm for Computational Psychiatry

This protocol update provides a revised overview of the normative modelling landscape, presents an updated standardized analytical protocol for neuroimaging data that incorporates federated and longitudinal approaches, offers practical guidance with open-source code via the refactored PCNtoolkit, and releases new community models for various imaging modalities to advance individual-level analysis in computational psychiatry.

Original authors: de Boer, A. A. A., Bayer, J. M. M., Fraza, C., Chavanne, A., Rehak Buckova, B., Tsilimparis, K., Serin, E., Bernas, A., Cirstian, R., Zabihi, M., Rutherford, S., Al Khaledi, A., Wolfers, T., Beckmann

Published 2026-02-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: de Boer, A. A. A., Bayer, J. M. M., Fraza, C., Chavanne, A., Rehak Buckova, B., Tsilimparis, K., Serin, E., Bernas, A., Cirstian, R., Zabihi, M., Rutherford, S., Al Khaledi, A., Wolfers, T., Beckmann, C., Marquand, A. F.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a doctor trying to figure out if a child is growing up healthy. In the past, you might have just compared that child to a single "average" height. But what if the child is naturally tall? Or what if they are from a family where everyone is short? A simple average doesn't tell the whole story.

This paper is essentially a new, high-tech instruction manual for building a "Growth Chart for the Brain."

Here is the breakdown of what the authors are doing, using simple analogies:

1. The Big Idea: "Brain Growth Charts"

Instead of just comparing a patient to a sick person or a healthy person (a simple "us vs. them" approach), this method builds a massive map of how healthy brains change as people age.

The Analogy: Think of a standard pediatric growth chart. It doesn't just say "average height." It shows a range: the 5th percentile (short), the 50th (average), and the 95th (tall). It tells you that a 10-year-old can be 4 feet tall or 5 feet tall and still be perfectly healthy.
The Application: This paper teaches scientists how to build that same kind of chart for the brain. It maps out how brain volume, thickness, and activity change from childhood to old age in healthy people. Then, when a patient comes in, the doctor can see exactly where their brain falls on that chart. Are they on the "tall" track? The "short" track? Or are they way off the chart entirely?

2. The Problem with the Old Way

Previously, researchers mostly did "Case-Control" studies.

The Old Way: They took a group of sick people and a group of healthy people and compared the averages.
The Flaw: It's like saying, "The average height of basketball players is 6'5", and the average height of jockeys is 5'2"." This tells you nothing about the 5'10" basketball player or the 6'0" jockey. It ignores the huge variety within groups.
The New Way: This paper moves us to Individual-Level Analysis. It asks, "Given this specific person's age, sex, and background, how does their specific brain compare to the healthy norm?"

3. The Tool: PCNtoolkit (The "Construction Kit")

The authors have updated a software package called PCNtoolkit. Think of this as a Lego construction kit for building these brain charts.

Version 1.0: They have completely rebuilt the kit to be easier to use.
New Features:
- Handling Messy Data: Real-world data is messy (missing pieces, weird outliers). The new kit has better tools to clean this up, like a filter that removes "bad apples" without throwing away the whole basket.
- Longitudinal Tracking: It can now track changes over time. Imagine watching a movie of a brain growing, rather than just looking at a single photo. It can tell you if a brain is "growing too fast" or "shrinking too fast" compared to the norm.
- Non-Gaussian Models: Sometimes brain data isn't a perfect bell curve (like height). Sometimes it's lopsided. The new kit can handle these weird shapes better.

4. The "Privacy Shield": Federated Learning

One of the biggest hurdles in medical research is privacy. Hospitals often can't share patient data because of strict laws.

The Analogy: Imagine five different chefs (hospitals) trying to create the perfect soup recipe. Usually, they would have to send their ingredients to one central kitchen to mix them. But what if they can't send the ingredients?
The Solution: This paper explains how to use Federated Learning. Instead of sending the ingredients (data), the chefs send their recipes (mathematical models) to a central spot. The central spot mixes the recipes to create a master recipe, then sends it back. The data never leaves the hospital, but the result is a model trained on data from all five hospitals. This paper shows how to do this with brain charts.

5. Why This Matters

This isn't just about math; it's about Precision Medicine.

Before: "You have schizophrenia." (A broad label).
After: "Your brain's volume in this specific area is at the 2nd percentile for your age and sex, which is a significant deviation from the healthy norm, suggesting a specific type of progression."

Summary

This paper is a protocol update. It's saying: "We have a new, better, and more flexible way to build 'Brain Growth Charts.' We have updated the software (PCNtoolkit) to make it easier to use, to handle privacy concerns, and to track changes over time. Here is the step-by-step guide on how to use it."

It transforms psychiatry from a game of "guessing the average" into a precise science of "measuring the individual."

1. Problem Statement

Traditional computational psychiatry has largely relied on case-control analysis, which assumes rigid categorical distinctions between healthy and clinical populations. This approach fails to capture the continuous and heterogeneous nature of psychiatric biology, often missing subtle individual deviations.

While Normative Modelling (NM) has emerged as a solution to map healthy developmental trajectories and quantify individual deviations, the field faces several challenges:

Methodological Fragmentation: Recent innovations (non-Gaussian models, longitudinal tracking, federated learning) have been implemented in disparate software tools.
Reproducibility and Usability: There is a need for a standardized, open-source protocol that integrates these advanced methods into a single, maintainable framework.
Data Privacy and Scale: Large-scale modeling requires aggregating data across sites, which is often hindered by privacy regulations and data sharing restrictions.

2. Methodology

The paper presents a comprehensive update to the Predictive Clinical Neuroscience toolkit (PCNtoolkit), specifically version 1.x.x, which serves as the primary software implementation for this protocol.

Core Framework

Paradigm: NM maps biological/behavioral response variables (e.g., cortical thickness) as a function of covariates (e.g., age, sex) and batch effects (e.g., scanner site) to estimate a population distribution. It outputs deviation scores (Z-scores) and centiles for individuals, indicating how far they deviate from the normative population.
Software Architecture: The toolkit is refactored into an object-oriented Python package. The central class is NormativeModel, which manages a collection of univariate RegressionModel objects (one per feature).

Key Algorithmic Components

Hierarchical Bayesian Regression (HBR): The default and most flexible modeling approach.
- Likelihoods: Supports Normal (Gaussian), SHASHb (for unbounded/asymmetric data), and Beta (for bounded data) likelihoods to handle non-Gaussian distributions.
- Priors: Uses non-centered hierarchical parameterizations to stabilize sampling. Priors can be fixed, random (batch effects), or linear/non-linear functions of covariates (using B-spline basis functions).
- Inference: Relies on Markov Chain Monte Carlo (MCMC) via PyMC, with convergence diagnostics (R-hat, ESS, MCSE) provided via ArviZ.
Data Handling:
- Preprocessing: Automated outlier detection (Z-score thresholding), missing value handling, and standardization.
- Batch Effects: Modeled as random effects to account for site/scanner variability, allowing for partial pooling across sites.
Advanced Workflows:
- Longitudinal NM: Moves beyond cross-sectional snapshots to track individual trajectories. It utilizes Z-diff scores (for two timepoints) and velocity modeling (for multiple timepoints) to detect "centile crossings" (significant deviations in developmental trajectory).
- Federated Learning: Enables model training across decentralized data without sharing raw data. Three strategies are supported:
  - Model Transfer: Applying a reference model's posterior as an informative prior to new local data.
  - Model Extension: Updating a reference model with new site data.
  - Model Merging: Combining independently fitted models.
Model Evaluation:
- Metrics: Includes mean-based metrics ( $R^2$ , RMSE) and probabilistic metrics crucial for NM: Mean Standardized Log Loss (MSLL) and Mean Absolute Centile Error (MACE).
- Diagnostics: Visual inspection of centile plots and Quantile-Quantile (QQ) plots to assess calibration.

3. Key Contributions

Standardized Protocol Update: Provides a revised, step-by-step analytical protocol (replacing the 2022 version) that integrates the latest methodological advancements into a unified workflow.
Refactored PCNtoolkit (v1.x.x): A major software overhaul offering:
- Support for non-Gaussian distributions (SHASHb, Beta).
- Native longitudinal modeling capabilities.
- Federated learning pipelines.
- Improved user interface and reproducibility features.
Open Resources:
- Release of updated lifespan reference models for cortical thickness, volumetric data, and diffusion-weighted imaging.
- Public code repository with Jupyter notebooks demonstrating the full workflow, from data selection to federated learning.
Community Engagement: Introduction of PCNPortal for collaborative model sharing and contribution guidelines for the open-source community.

4. Results

The paper validates the protocol using a multi-site dataset (FCON1000 and OpenNeuro clinical datasets) involving cortical thickness data:

Model Convergence: Hierarchical Bayesian models demonstrated robust convergence with potential scale reduction factors ( $\hat{R}$ ) $\le 1.01$ and effective sample sizes (ESS) > 1000 across all parameters.
Calibration: Visual diagnostics (centile and QQ plots) confirmed that the models accurately captured the conditional distribution of the data, including heteroskedasticity (variance changing with age).
Performance: The models successfully estimated deviation scores independent of covariates (age, sex, site), demonstrating the ability to isolate biologically relevant variance from technical and demographic noise.
Scalability: The protocol successfully demonstrated parallelized fitting on HPC clusters (Slurm) and the feasibility of federated learning strategies.

5. Significance

Shift to Precision Psychiatry: This protocol facilitates the transition from group-level statistics to individual-level inference, allowing clinicians and researchers to quantify how a specific patient deviates from a healthy normative trajectory.
Handling Real-World Complexity: By supporting non-Gaussian data, longitudinal tracking, and federated learning, the toolkit addresses the practical limitations of real-world clinical data (e.g., skewed distributions, multi-site heterogeneity, and privacy constraints).
Reproducibility: The open-source nature of PCNtoolkit and the detailed protocol ensure that normative modeling can be replicated and extended by the broader scientific community, fostering the development of reliable biomarkers for psychiatric disorders.
Future-Proofing: The modular, object-oriented design allows for the easy integration of future algorithmic innovations, ensuring the framework remains relevant as computational methods evolve.