Human brain normative modeling by brain eigenmodes

1274 

Abstract Submission 

Sina Mansour L.¹, B. T. Thomas Yeo², Maria Di Biase³, Andrew Zalesky⁴

¹National University of Singapore, Singapore, NA, ²National University of Singapore, Singapore, Singapore, ³The University of Melbourne, Parkville, Victoria, ⁴The University of Melbourne, Melbourne, Victoria

Sina Mansour L.  
National University of Singapore
Singapore, NA

B. T. Thomas Yeo  
National University of Singapore
Singapore, Singapore

Maria Di Biase  
The University of Melbourne
Parkville, Victoria

Andrew ZALESKY, PhD  
The University of Melbourne
Melbourne, Victoria

Normative models (NMs) construct reference charts for population-wide distributions of a biological phenotype [1]. Analogous to growth charts in medicine, assessing a child's height relative to their age and sex, normative brain charting is a framework for modeling variations in structural brain phenotypes, such as cortical thickness. Previous research has demonstrated NM's efficacy in accurately capturing the heterogeneity of normative deviations in brain structure [2]. Normative charting of brain MRI data holds the potential to yield insightful spatial estimates of variations in cortical phenotypes [3]. Nevertheless, methodological limitations have thus far impeded the development of NMs with high spatial precision. Recent advances have increased the spatial resolution of NMs to predefined brain atlases [4]. Despite these notable efforts, existing approaches fall short of achieving spatially detailed norms comparable to the original resolution of MRI data. The establishment of a normative framework to detect subtle spatial nuances remains an unfulfilled aspiration for precision psychiatry.

To establish and evaluate our normative reference models, we combined data from three distinct Human Connectome Project (HCP) cohorts to cover the human lifespan (HCP Development [5], Young Adult [6], and Aging [7]), comprising 2,473 individuals (54.7% female) aged 5 to 100. Cortical thickness was extracted for each scan using HCP's minimal preprocessing pipeline [8].

Computing high-resolution NMs is challenging due to the high dimensionality of the feature space. Appropriate low-dimensional encoding of cortical phenotypes could hence enable the development of computationally tractable high-resolution normative references. To this end, we utilized brain eigenmodes [9] as basis functions for information reconstruction (Fig. 1a,b). We constructed a sparse, high-resolution connectome and used graph signal processing techniques to generate connectome eigenmodes via singular value decomposition of a random-walk Laplacian shift operator.

These eigenmodes were utilized as a basis for normative reconstruction by graph signal filtering [10]. Specifically, 2,000 brain eigenmodes encoded the high-resolution thickness information in the graph frequency domain. Thickness loading on each eigenmode formed a distinct spectral phenotype of the cohort. Distinct hierarchical Bayesian regression NMs were trained to model spectral phenotypes as a function of age and sex while accounting for scanner/site effects. This facilitated the generalization of pre-trained reference NMs to unseen spatial normative queries (Fig. 1c,d).

After training spectral NMs, we conducted a comprehensive evaluation of their ability to infer norms in comparison to a direct model. Posterior predictive distributions of spectral NMs were used to reconstruct normative ranges for alternative families of spatial queries. Notably, global spatial queries (e.g., mean thickness), regional spatial queries (e.g., thickness of a functional network), and high-resolution queries focused on specific cortical vertices were tested to evaluate goodness of fit (Fig. 4). Our findings indicate that spectral NMs are capable of generating effective normative ranges at various spatial resolutions.

We present a novel approach to extract versatile normative ranges from a set of pre-trained NMs constructed from brain eigenmodes. Our findings underscore the effectiveness of spectral NMs in generating accurate normative ranges. This obviates the necessity for the computationally intensive task of fitting new reference models for distinct spatial queries. This can inform clinical studies investigating normative trajectories of healthy brain development and aging, all without the need for direct access to or extensive processing of large imaging datasets.

Lifespan Development Other ¹

Bayesian Modeling

Exploratory Modeling and Artifact Removal

Methods Development

Anatomical MRI ²

MRI

NORMAL HUMAN

Open-Source Code

Other - Normative modeling