Normative discrimination for stratification - proof of concept in the European autism project

1898 

Abstract Submission 

Tristan Looden¹, Alberto Llera², Dorothea Floris³, Jan Buitelaar², Christian Beckmann²

¹Donders Institute, Nijmegen, Gelderland, ²Donders Institute for Brain, Cognition, and Behavior, Radboud University Nijmegen, Nijmegen, NL, Nijmegen, Netherlands, ³Methods of Plasticity Research, Department of Psychology, University of Zürich, Zürich, Switzerland

Tristan Looden  
Donders Institute
Nijmegen, Gelderland

Alberto Llera  
Donders Institute for Brain, Cognition, and Behavior, Radboud University Nijmegen, Nijmegen, NL
Nijmegen, Netherlands

Dorothea Floris  
Methods of Plasticity Research, Department of Psychology, University of Zürich
Zürich, Switzerland

Jan Buitelaar  
Donders Institute for Brain, Cognition, and Behavior, Radboud University Nijmegen, Nijmegen, NL
Nijmegen, Netherlands

Christian Beckmann  
Donders Institute for Brain, Cognition, and Behavior, Radboud University Nijmegen, Nijmegen, NL
Nijmegen, Netherlands

Autism spectrum disorder is a complex neurodevelopmental condition (American Psychiatric Association 2013). Prior work in autism has shown significant multivariate relationships between brain functional connectivity atypicality relative to typically developing controls and cognitive autism-associated measures across different tasks (Looden et al. 2022). The fact that we can identify a relationship between brain functional atypicality and autism-associated measures makes it an interesting target for subtyping. In this project, we investigate whether such subtyping of autistic individuals on the basis of functional connectivity also explains relevant variance in autism measurements. As a crucial methodological step for dimensionality reduction from covariance space to clusters we will introduce and validate Spatial Patterns for Discriminative Estimation (SPADE) an implementation of the Fukunaga-Koontz transform (Llera, A., Chauvin R., Mulders P., Naaijen J., Mennes M. 2019). SPADE finds linear filters in high-dimensional covariance space that optimally separate classes or observations, and furthermore provides visually interpretable spatial maps. We develop an individualized implementation of this framework to characterize individuals with respect to a reference cohort, on the basis of which further analysis is made possible.

All analyses were based on subsamples from the EU-AIMS/AIMS2TRIALS multisite Longitudinal European Autism Project (LEAP) with participants between 6 and 30 years of age (Charman et al. 2017). This study includes autistic (N=282) and typically developing (N=221) participants and uses data drawn from five different fMRI task paradigms, each sampling brain functional connectivity from different cognitive brain states. The following analysis is done in each of the five tasks. First, SPADE is used on task data to find a single spatial discriminative filter that best separate each autistic individual from the typically developing group (see fig 1). This subject-specific characterization of autistic individuals is then used to form a similarity matrix which is fed into a spectral decomposition to uncover subtype characteristics in the similarity space. The optimal number of clusters in the two-dimensional spectral output is estimated by minimizing the Bayesian Information Criterion (BIC). The resulting clustering solutions are subsequently used in a linear model to explain behaviour in a well-rounded selection of autism-associated measures (corrected for age and sex). We report the variance explained by these clusters in each clinical measure as well as the model F-test for overall significance per measure. Finally, we use the spatial filters learned by SPADE which formed the basis of the clustering to inform us of the most relevant brain regions for each cluster.

We found a significant relationship between the clustering solution (5 clusters) and one or more clinical autism measures in each of the five tasks (FDR corrected per task). Variance explained for each measure ranges between 7-11% in the significant relationships. Specifically, the cognitive demand induced by the Hariri task formed a fruitful source of covariances on the basis of which clinically relevant clustering could be performed. See figure 2 for more detail in the Hariri task.

We presented a proof of concept towards a new normative modeling approach for population fmri data with respect to a reference cohort. The SPADE for stratification model uses maximal discrimination filters to find differences from each subject to the reference cohort and we show preliminary results that the process yields behaviouraly relevant stratifications of autistic participants. The clinically validated covariance-based subclusters in autism may furthermore point to groups in autism with differential etiologies as targets for future research.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ²

Classification and Predictive Modeling

Connectivity (eg. functional, effective, structural)

Methods Development ¹

Multivariate Approaches

Autism

Data analysis

Machine Learning

Modeling

MRI

Multivariate

Statistical Methods