Prediction of longitudinal anxiety in adolescents using mixed effects random forest regression

606 

Abstract Submission 

Paola Odriozola¹, Amanda Baker¹, Claire Waller¹, Nancy Le¹, Katie Bessette¹, Lucina Uddin², Tara Peris¹, Adriana Galván¹

¹UCLA, Los Angeles, CA, ²Department of Psychiatry and Biobehavioral Sciences, University of California, Los Angeles, Los Angeles, CA

Paola Odriozola, Ph.D.  
UCLA
Los Angeles, CA

Amanda Baker, Ph.D.  
UCLA
Los Angeles, CA

Claire Waller  
UCLA
Los Angeles, CA

Nancy Le  
UCLA
Los Angeles, CA

Katie Bessette, Ph.D.  
UCLA
Los Angeles, CA

Lucina Uddin  
Department of Psychiatry and Biobehavioral Sciences, University of California, Los Angeles
Los Angeles, CA

Tara Peris, Ph.D.  
UCLA
Los Angeles, CA

Adriana Galván, Ph.D.  
UCLA
Los Angeles, CA

Adolescence is a peak time for the onset of psychiatric disorders, with anxiety disorders being the most common, affecting as many as 1 in 3 youths (Beesdo et al., 2009; Kessler et al., 2005; Merikangas & Swanson, 2010). Despite its significant costs, understanding of factors that shape the persistence/remittance of anxiety over time remains limited. Using machine learning methods with multivariate longitudinal behavioral, clinical, and fMRI data from early to mid-adolescents, we took a data-driven approach to investigate whether it was possible to predict whose anxiety will worsen, remain the same, or remit years later. We hypothesized that mixed effects random forest regression would enable prediction of anxiety outcomes with high precision, and that the functional connectivity of brain regions previously shown to be implicated in anxiety (e.g., amygdala, hippocampus, ventral striatum, anterior insula (AI), anterior cingulate cortex (ACC), and ventromedial prefrontal cortex(vmPFC)) would be of highest importance in predicting anxiety outcomes.

132 adolescent participants (61 F : 71 M; 11.4 +/-1.5 years at time 1) completed the Development of Anxiety in Youth Study (Galván & Peris, 2020), a prospective longitudinal study that occurred annually for 3 years. Participants completed a resting state fMRI scan, the Screen for Child Anxiety Related Disorders (SCARED) child and parent report questionnaires (Birmaher et al., 1997), and demographic questionnaires at each visit. We used the AAL3 atlas (Rolls et al., 2020) to parcellate the brain, and computed the functional connectivity between each parcel to generate a correlation matrix using AFNI FATCAT (Taylor & Saad, 2013). We then submitted scaled demographic, behavioral, and functional connectivity data to a stochastic mixed effects random forest regression analysis (sMERF) implemented in R using the LongituRF package (Capitaine et al., 2021). This package combines the feature selection aspects of random forests with an extension to include mixed-effects models to account for repeated measures for high-dimensional longitudinal data. We used a standard Ornstein-Uhlenbeck process which allows the covariance structure to vary over time (Capitaine et al., 2021). We used 80% of the data for training, and the other 20% for testing the model. The model contained 13,538 predictors which included functional connectivity values (functional connectivity of all AAL3 parcels), and demographic variables (i.e., age, sex at birth, race, ethnicity, family income, IQ, etc.) and the outcome of interest was child-reported SCARED total score. Prediction errors were calculated as root mean square error with 25 training/test set random splits.

Prediction of future anxiety symptoms using sMERF yielded a root mean square error of 0.966. The top 5 variables that yielded the highest importance in the random forest model included (in order of relative importance): functional connectivity of the left gyrus rectus to left nucleus accumbens (Nacc), right crus I of cerebellar hemisphere to left locus coeruleus, right crus I of cerebellar hemisphere to left pregenual ACC, left temporal pole (middle temporal gyrus) to right Nacc, and lobule I and II of vermis to left lateral posterior thalamus.

Anxiety disorders often emerge during childhood and adolescence, yet not all youth benefit sufficiently from current evidence-based treatments and long-term outcomes are variable (Bai et al., 2023). Results from the present study suggest that resting functional connectivity between regions regions often overlooked in studies of anxiety- such as the Nacc, ACC, and cerebellum- may play a larger role in predicting anxiety outcomes. Increasing our understanding of factors that predict future anxiety outcomes across development is crucial for identifying optimal windows for prevention or intervention and specifying new targets for intervention for youth struggling with anxiety.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Early life, Adolescence, Aging

Classification and Predictive Modeling ²

Task-Independent and Resting-State Analysis

BOLD fMRI

Anxiety

Development

FUNCTIONAL MRI

Machine Learning

Multivariate

Other - longitudinal