Subcortical emotion-related circuitry serves as brain-based classifier of math anxiety in children

720 

Abstract Submission 

Yi-Cheng Cho¹, Chan-Tat Ng¹, Ting-Ting Chang^1,2

¹Department of Psychology, National Chengchi University, Taipei City, Taiwan, ²Research Center for Mind, Brain, and Learning, National Chengchi University, Taipei City, Taiwan

Yi-Cheng Cho  
Department of Psychology, National Chengchi University
Taipei City, Taiwan

Chan-Tat Ng  
Department of Psychology, National Chengchi University
Taipei City, Taiwan

Ting-Ting Chang  
Department of Psychology, National Chengchi University|Research Center for Mind, Brain, and Learning, National Chengchi University
Taipei City, Taiwan|Taipei City, Taiwan

Math anxiety (MA) has been shown to severely impact children's mathematical performance. While prior research identified brain regions associated of math and emotion like intraparietal sulcus, amygdala, and insula (Lyons and Beilock, 2012, Young et al., 2012, Pletzer et al., 2015), understanding of the functional circuits underlying MA remains scarce. This study leveraged data-driven machine learning to fully capture the potential whole-brain resting-state functional connectivity (FC) features associated with MA.

We recruited 133 participants aged 7 to 19 years (M = 10.40, SD = 0.45; 76 females and 57 males) in this study, categorized as high and low MA based on mean MA scores from the Chinese-adapted Child Math Anxiety Questionnaire (CMAQ; Ng et al., 2022). Participants underwent a series of cognitive assessments, including Block Design, Vocabulary, and Digit Span subtests from the WISC-IV, along with an arithmetic fluency test (Chang et al., 2018) (Fig. 1A). During the MRI session, participants underwent an eye-closed resting-state functional scans (Fig. 1B). The functional imaging data were preprocessed using SPM12, and whole-brain resting-state FC measures were computed across 166 brain regions of interest (ROIs) based on the automated anatomical labeled (AAL3) atlas, resulting in 13,695 ROI-to-ROI FC measures (Fig. 1C). Feature selection was then conducted in two stages: initially through Pearson correlation to identify FC measures that were linearly correlated to MA (with a significance threshold of p < .001), followed by mutual information techniques (Vergara & Estévez, 2014) for feature refinement, narrowing to the top 40 most informative features. Two distinct features sets were established: brain-only features comprised only of FC measures and brain-and-cognition features which included both FC measures and cognitive scores. These features were then used respectively in a random-forest machine learning model with 5-fold cross validation for MA classification. The model trained on 80% of the data, reserving 20% for validation and feature importance analyses (Fig. 1D).

The initial stage of feature selection found 197 FC measures significantly correlating with MA. These measures were primarily associated with connections within or between subcortical regions (94.9% of the significant connections), predominately the thalamus and cerebellum. Following the application of mutual information techniques, we distilled these measures down to two optimized sets of 40 features each for the subsequent MA classification. Both brain-only and brain-and-cognition models demonstrated high MA classification accuracy scores (81.5% and 85.2%, respectively) (Fig. 2). Further feature importance analyses highlighted the pivotal role of subcortical connections, especially those involving the thalamus (20 out of 40 features) and cerebellum (12 out of 40 features). Additionally, we identified connections between subcortical and cortical regions, especially those implicated in emotional processing (such as the anterior cingulate cortex and insula) and the default mode network (including the angular gyrus and posterior cingulate cortex, mean |SHAP| value exceeds 0.1). While the inclusion of cognitive abilities in the brain-and-cognition model, notably the subtraction component of arithmetic fluency, enhanced model accuracy, the predominant contributors to the classification success were connectivity measures.

Using a data-driven machine learning approach, our study highlights the pivotal contributions of subcortical regions for MA. Particularly, we demonstrated that subcortical connections involving thalamus and cerebellum, rather than within-cortical connection or cognitive ability, were among the top contributors in MA classification. This integrated approach deepens our understanding of math anxiety's multifaceted nature, providing potential subcortical brain-based biomarkers for MA identification and remediation.

Emotion and Motivation Other ¹

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

Anxiety

FUNCTIONAL MRI

Modeling

Other - Math Anxiety; Machine Learning