A principled approach to predict functional connectivity from age and sex for normative modelling

1785 

Abstract Submission 

Janus Kobbersmed¹, Diego Vidaurre¹

¹Aarhus University, Aarhus, Aarhus C

Janus Kobbersmed  
Aarhus University
Aarhus, Aarhus C

Diego Vidaurre  
Aarhus University
Aarhus, Aarhus C

Many neurological and psychiatric diseases lack measures of brain functioning as indicators of brain pathology. The search for such biomarkers is complicated by the fact that many brain diseases and disorders, like e.g. autism or ADHD, affect multiple brain regions and their interplay. Furthermore, these conditions are often described as a spectrum from normal to abnormal rather than a sick-healthy dichotomy (Marquand et al. 2016). Recently, normative modelling of brain structure and function has gained increasing popularity (Rutherford et al. 2023). This approach seeks to characterize the distributions of population brain measures to identify the range of normal brain measures. Normative modelling of brain function requires data from a large population of healthy subjects as well as a method to predict normal brain functioning from sex and age. In fMRI studies, brain functioning is often assessed via functional connectivity (FC) which is the matrix of covariances between brain regions (Smith et al. 2013). However, predicting FC from sex and age is challenging because of the mathematical structure and high dimensionality of the FC matrix (Hoff and Niu 2012). Current normative modelling studies have mainly focused on predicting the individual elements of the FC matrix rather than the FC matrix as a whole and its structure (e.g. (Looden et al. 2022)). To address this gap, we investigate a newly developed method for covariance matrix regression in a normative modelling setting. Using resting fMRI data from the Human Connectome Project (HCP), we adapt this method to predict FC from sex and age for characterization of normal brain functioning.

FC was derived from a set of resting fMRI scans from 1000 subjects from the HCP. The parcellation was an overlapping, spatial parcellation with 50 ICA parcels. In order to predict FC from sex and age, we used a newly developed method called covariate-assisted principal regression (CAPR), which was found in literature (Zhao et al. 2021). The method finds a transformation of the subjects' covariance matrices by identifying projections that depend on chosen covariates. We first assessed the consistency and scalability of the method and how it depended on the parcellation, covariates, and number of subjects and timepoints. We then adapted the method to our normative modelling framework and identified 2 age-dependent projections for each sex category in a training set of 800 healthy subjects. This allowed us to find the distribution of the transformed FC in the training set. Each of 200 healthy out-of-sample test subjects were then assigned a Z-score of their FC given sex and age. The whole process was repeated using 5-fold cross-validation to validate the model's capability of defining normal functioning in a healthy population.

Having established the consistency of the CAPR method, we adapted it for use in normative modelling and identified relevant transformations of the FC data in the HCP dataset. This reduced dimensionality of the data to summarize individual FC in 2 age-dependent derived values (Figure 1). The prediction of the derived brain measures allowed assigning Z-scores for each of 200 out-of-sample test subjects (4 examples shown in Figure 1). The distribution of the Z-scores obtained from cross-validation showed that we were able to characterize brain functioning in the healthy population. Bootstrapping revealed that this method was computationally scalable to larger datasets (>20000 subjects).

Taking advantage of a large, publicly available brain dataset, this work suggests a useful method to extract relevant age- and sex-related characteristics in FC and use these to predict normal brain functioning. The method successfully characterizes the FC in a healthy population and is scalable. This paves the way for normative modeling using larger datasets as well as identification of brain functioning pathology.

Classification and Predictive Modeling ²

fMRI Connectivity and Network Modeling ¹

FUNCTIONAL MRI

Modeling

NORMAL HUMAN

Other - Normative modelling