Using Machine Learning to Identify Age- and Domain-Specific Biobehavioral Risk Factors for Callous-Unemotional Traits Across Infancy, Toddlerhood, and Early Childhood.

• In Event: Heterogeneity in Externalizing Psychopathology: the Contribution of Multiple Etiological Pathways Across Childhood and Adolescence.

Abstract

Callous-unemotional (CU) traits are defined by low empathy and guilt, and reduced sensitivity to others’ emotions. CU traits can be identified in early childhood and predict risk for disruptive behavior disorders (DBDs), even in community samples. Standard treatments for DBDs are less effective for children with CU traits, highlighting an urgent need to identify specific and modifiable risk factors for CU traits. Although research has established endogenous (e.g., temperament and physiological sensitivity) and exogenous (e.g., parenting practices and exposure to violence) risk factors for CU traits, no prior studies have jointly tested different endogenous and exogenous risk factors assessed multiple times at different ages within a single framework. Thus, we do not know which exogenous and endogenous factors constitute the most risk for CU traits and when they matter. The current study addresses this knowledge gap by leveraging supervised machine learning algorithms to identify the specific endogenous and exogenous factors that best signal risk for CU traits across infancy, toddlerhood, and middle childhood.
Using a prospective longitudinal sample of 208 children recruited at birth, we analyzed the relative importance of endogenous (e.g., autonomic nervous system functioning and temperament) and exogenous (e.g., observed parenting) risk factors assessed in infancy (i.e., 3-12 months), toddlerhood (i.e., 18-30 months), and early childhood (i.e., 36 months) for predicting risk for CU traits in late childhood (84 months). Models predicting severity of CU traits using Random Forest with boosting (R package: gbm, Greenwall et al., 2022) revealed that in early childhood, parent-reported soothability, observed parental engagement, and baseline cortisol in children had the highest relative predictive importance. Baseline cortisol in both toddlerhood and infancy also emerged as important predictors of CU traits, as did lower observed maternal sensitivity and parent-reported soothability in infancy. Random Forest classification models (R packages: randomForest, Liaw & Wiener, 2002; gbm, Greenwall et al., 2022) based on previously established diagnostic thresholds predicted clinical levels of CU traits with 85% specificity (AUC = .79) and indicated that in early childhood, soothability, and baseline vagal tone and cortisol held the highest predictive importance. Observed parental engagement, positive regard, and negative regard in infancy and toddlerhood, as well as parent-reported fear in infancy also influenced the prediction of CU traits. For the first time, findings that leverage a machine learning approach have isolated age- and domain-specific risk factors for CU traits. These findings lay the foundation for future research that applies iterative statistical methods to approach robust longitudinal secondary data analysis and informs targeted, domain- and age-specific interventions to personalize our treatments for DBD and CU traits.

Authors

• Deaven Winebrake, Boston University
  Presenting Author

• Anthony J. Rosellini

• Cathi Propper, University of North Carolina at Chapel Hill

• Rebecca Waller, University of Pennsylvania

• Nicholas James Wagner, Boston University