A Comparative Machine Learning Study of Connectivity-Based Biomarkers of Schizophrenia

497 

Abstract Submission 

Victoria Shevchenko¹, R. Austin Benn¹, Robert Scholz^1,2,3, Wei Wei¹, Carla Pallavicini¹, Ulysse Klatzmann¹, Francesco Alberti¹, Theodore Satterthwaite⁴, Demian Wassermann⁵, Pierre-Louis Bazin⁶, Daniel Margulies¹

¹Université Paris Cité, INCC UMR 8002, CNRS, Paris, France, ²Max Planck School of Cognition, Leipzig, Germany, ³Wilhelm Wundt Institute for Psychology, Leipzig University, Leipzig, Germany, ⁴University of Pennsylvania, Perelman School of Medicine, Philadelphia, PA, USA, ⁵Université Paris-Saclay, Inria, CEA, Paris, France, ⁶Full brain picture Analytics, Leiden, The Netherlands

Victoria Shevchenko  
Université Paris Cité, INCC UMR 8002, CNRS
Paris, France

R. Austin Benn  
Université Paris Cité, INCC UMR 8002, CNRS
Paris, France

Robert Scholz  
Université Paris Cité, INCC UMR 8002, CNRS|Max Planck School of Cognition|Wilhelm Wundt Institute for Psychology, Leipzig University
Paris, France|Leipzig, Germany|Leipzig, Germany

Wei Wei  
Université Paris Cité, INCC UMR 8002, CNRS
Paris, France

Carla Pallavicini  
Université Paris Cité, INCC UMR 8002, CNRS
Paris, France

Ulysse Klatzmann  
Université Paris Cité, INCC UMR 8002, CNRS
Paris, France

Francesco Alberti  
Université Paris Cité, INCC UMR 8002, CNRS
Paris, France

Theodore Satterthwaite  
University of Pennsylvania, Perelman School of Medicine
Philadelphia, PA, USA

Demian Wassermann  
Université Paris-Saclay, Inria, CEA
Paris, France

Pierre-Louis Bazin  
Full brain picture Analytics
Leiden, The Netherlands

Daniel Margulies  
Université Paris Cité, INCC UMR 8002, CNRS
Paris, France

Functional connectivity holds promise as a biomarker of psychiatric disorders. Yet, its high dimensionality, combined with small sample sizes in clinical research, increases the risk of overfitting when the aim is prediction (Serin et al., 2021). Recently, low-dimensional representations of the connectome such as cortical gradients (Jung et al., 2022; Margulies et al., 2016) and gradient dispersion (Bethlehem et al., 2020) have been proposed, with studies noting consistent gradient and dispersion differences in psychiatric conditions (Dong et al., 2021; Pasquini et al., 2022; Xia et al., 2022). However, it is unknown which of these derivatives has the highest predictive capacity and how they compare to raw connectivity. Our study evaluates which connectome features -- functional connectivity, gradients, or gradient dispersion -- best predict schizophrenia (Figure 1A) and analyzes the impact of feature quantity on model selection.

We compiled resting state fMRI data from three open-source datasets: COBRE (Aine et al., 2017), UCLA Consortium for Neuropsychiatric Phenomics (Poldrack et al., 2016) and SRPBS multidisorder MRI dataset (Tanaka et al., 2021), totaling 996 individuals. After excluding 40 subjects due to motion artifacts (FD > 0.5mm), the sample included 248 patients with schizophrenia and 688 controls.

Preprocessing of MRI data was done with fMRIPrep 20.2.1 (Esteban et al., 2019) with BOLD timeseries parcellated using Schaefer parcellation (1000 parcels). We correlated (Pearson) region-wise timeseries to produce connectivity matrices. To compute cortical gradients, we z-transformed and thresholded (10%) the matrices and applied principal component analysis (PCA). Next, we computed gradient dispersion using two distinct approaches: 1. Bethlehem et al. (2020) and 2. K-Nearest Neighbors. We refer to these two types of dispersion as centroid and neighborhood dispersion, respectively.

Next, for each subject, we concatenated the flattened matrices, 200 gradients and centroid and neighborhood dispersion in one array. We applied PCA to each feature type separately across subjects. To ensure that more than 1 component is extracted from each feature type, we retained 20% of variance except for the centroid dispersion for which all variance was included. We fitted L2-regularized logistic regression on the training set and computed permutation component importance (Figure 1C, see caption). Finally, we inverse transformed component importance for each feature type to obtain feature importance which was used to select 100-10000 features per feature type for further assessments of model performance. Using Pycaret, we fit 13 models on the subsets (for the list of models, see Figure 2D). The analytic workflow used to compute component and feature permutation importance is depicted in Figure 1.

Our analyses yielded unexpected results. Out of all feature types tested in this study, connectivity had the largest component permutation importance (Figure 2A) and consistently showed the best predictive performance (Figure 2B). Connectivity also performed better across different numbers of features selected based on the importance in the feature space (Figure 2C). In addition, as the number of features increased, linear models tended to outperform more complex models.

The emergence of novel connectivity-based methods broadens our toolkit for predicting psychiatric disorders, but it also necessitates empirical testing. Surprisingly, our findings indicate that functional connectivity outperforms its more recent, low-dimensional derivatives such as cortical gradients and gradient dispersion in predicting schizophrenia. This suggests that the information within the connectome's specific edges is crucial for distinguishing between neurotypical individuals and those with schizophrenia. Further transdiagnostic studies are necessary to establish whether this tendency is consistent across the psychiatric spectrum.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

Task-Independent and Resting-State Analysis

Workflows

Computational Neuroscience

Cortex

FUNCTIONAL MRI

Modeling

Psychiatric Disorders

Schizophrenia

Workflows