A normative reference framework for resting-state network connectivity in psychiatry

Current symptom-based diagnostic systems group patients with highly heterogeneous clinical presentations and lack biomarkers to guide treatment selection. Precision psychiatry requires moving beyond traditional case-control approaches - which assume homogeneity within diagnostic categories - toward frameworks that quantify individual dysfunction. This shift enables mapping relationships between symptoms, behavior, and biology across diagnostic boundaries.

Normative modeling (NM) provides a statistical framework to define the typical distribution of brain measures in a healthy population as a function of relevant covariates (e.g., age, sex). Rather than comparing group averages, NM quantifies individual deviations from normative expectations, yielding subject-level deviation scores that capture heterogeneity obscured by group statistics. These deviation profiles can be linked to symptoms, behavior, and treatment response. Resting-state functional connectivity (rs-FC), reflecting large-scale brain organization and known to be altered across psychiatric conditions, represents a promising biomarker.

We developed and validated NMs of rs-FC across the seven canonical networks defined by Yeo et al. NMs were trained on a large multi-site healthy cohort (N = 4,818) from 22 studies (mean age = 33,5 ± 22,6 years; range 5–95 years; 43% female). FC was reconstructed using CATO, Pearson correlations were computed between mean time series of Schaefer-100 parcels and subsequently aggregated into seven network-level values representing average within-network connectivity. NMs were implemented using the Predictive Clinical Neuroscience Toolkit with Hierarchical Bayesian Regression. The reference cohort was split into 80% training and 20% test sets, stratified by age, sex, and site. NMs revealed robust age-dependent trajectories across all seven networks, demonstrating strong performance in both training (mean R² = 0.24, range: 0.14-0.35) and held-out test data (mean R² = 0.21, range: 0.11-0.29), confirming strong generalizability. The resulting framework provides an interpretable, reusable reference that enables subject-specific deviation mapping in independent psychiatric datasets.

The pretrained NMs are being applied to an independent transdiagnostic dataset (N = 2,250: 1,017 healthy controls, 1,335 patients; mean age = 35,7 ± 13,1 years; age range = 18 - 65 years; 63% female) comprising individuals with major depressive disorder (n = 956), schizophrenia spectrum disorders (n = 131), and bipolar disorder (n = 136). The dataset is stratified by age, sex, and site into healthy adaptation (n = 200) and test (n = 817) subsets for model calibration, with remaining patients (n = 1,233) used to generate individualized deviation scores. Ongoing multidimensional phenotypic profiling will determine whether individuals exhibiting similar rs-FC deviation patterns share common characteristics across diagnosis, symptoms, biomarkers, genetics, and environmental factors - potentially revealing transdiagnostic subtypes invisible to symptom-based classification.

This work establishes the first large-scale normative framework for resting-state network connectivity in the context of psychiatric disorders. The framework may enable: (1) individual-level quantification of functional connectome atypicality, (2) multidimensional phenotyping of patients with shared network dysfunction, and (3) stratification for targeted interventions based on biological signatures. The pretrained models will be made openly available, providing an immediately usable tool for precision psychiatry applications.

References

No references