Spectral Normative Modeling (SNM)

This repository hosts the scripts and code used for our manuscript on Spectral Normative Modeling of Brain Structure. The study focuses on normative modeling of brain development using spectral methods, with applications to both healthy and clinical populations.

---

Overview

The repository is organized into 7 sets of notebooks, each corresponding to a specific step in the analytical pipeline. Below is a summary of what each step accomplishes:

1. Data Import Scripts (code/notebooks/01_data_import)

• Process and clean cortical thickness data from healthy participants in the HCP lifespan datasets.
• Prepare data for downstream analysis.

2. Data Aggregation Scripts (code/notebooks/02_data_aggregation)

• Combine cleaned data from multiple datasets.
• Randomly split data into training and test sets for model validation.

3. Normative Models (code/notebooks/03_normative_models)

• Implement a basic normative model architecture using Hierarchical Bayesian Regression (via the PyMC Python package).
• Normative model for a single variable (e.g., mean cortical thickness) as a function of covariates (age, sex, site).
• This architecture is used for both direct and spectral implementations.

4. Spectral Basis Construction (code/notebooks/04_normative_kernels)

• Construct spectral kernels/basis functions to encode high-resolution cortical phenotypes.
• Map connectome-based brain eigenmodes via singular value decomposition of a random walk graph Laplacian shift operator (see notebook 04_04_02).

5. Spectral Normative Model Fitting (code/05_kernel_normative_models)

• Fit a prototype of the Spectral Normative Model (SNM).
• Verify the generation of normative ranges (see notebook 05_01_05).

6. Performance Evaluation Scripts (code/06_performance_evaluation)

• Evaluate the accuracy of SNMs with varying numbers of modes.
• Compare SNM performance to a direct model.
• Generate comparative figures presented in the manuscript.

7. Clinical (AD) Evaluations (code/07_clinical_evaluation)

• Clean and preprocess data from a clinical Alzheimer's Disease (AD) sample (MACC Harmonization dataset).
• Fine-tune the healthy SNM to the new site by learning harmonization parameters.
• Perform clinical evaluations and reproduce figures reported in the manuscript.

---

Citation

If you use this repository in your research, please cite our manuscript:

Mansour L., S., et al. (2025). Spectral Normative Modeling of Brain Structure. medRxiv. DOI: 10.1101/2025.01.16.25320639

---

Contact:

If you have any questions or need assistance, please don't hesitate to reach out.