Developmental Changes in Learning to Predict Others' Preferences: Implications for Autism Spectrum Disorder

Background:

Many of our efforts in social interactions are dedicated to optimizing our predictions of others’ preferences, mental states, and behaviors, an ability referred to as Theory of Mind (ToM). Research over the past two decades has established that children with autism spectrum disorder (ASD) show significant developmental delays in ToM. . In consequence, behavioral interventions for ASD focus mostly on improving ToM skills. However, a mechanistic understanding of how we improve our predictions of others’ mental states over time, and an understanding of developmental changes in these learning processes are lacking. 

 Objectives:  

We aim to gain a mechanistic understanding of how humans optimize predictions of others’ preferences over the course of typical development. Optimizing predictions is defined as reducing the difference between expected and actual outcomes, i.e., prediction errors. In a next step, we will investigate developmental differences in encoding and updating social prediction errors between typically developed adolescents and adolescents with ASD. 

Methods:  

We utilize a novel preference task to investigate the behavioral and neural mechanisms involved in updating social predictions and learning from prediction errors. To ensure high ecological validity, the task involves real social feedback. Typically developed adults (N=21) and adolescents (N=23) are asked to predict the preferences of three different people for a number of items (food types, beauty products and leisure activities). They subsequently receive trial-by-trial feedback about others’ actual preference ratings. After completing the task, participants rate their own preferences for the same items (see Figure 1 A). We use reinforcement learning (RL) models to describe participants’ changes in ratings over time. On the neural level, we model participant’s trial-by-trial prediction errors (the absolute difference between participant’s rating of the other person’s preference and the person’s actual preference rating) and the perceived self-other difference (the absolute difference between participant’s ratings of their own preference versus that of the other person). We also used estimated prediction errors and estimated ratings from the winning RL model to predict trial-by-trial changes in brain activity.

Results:  

Estimating others’ preferences relies on two components: reinforcement learning based on past feedback and participants’ own preferences for an item. Learning processes differ across development. Adolescents are slower at updating their predictions based on past feedback (see Figure 1 B); older adolescents show more updating (i.e., smaller prediction errors). Paralleling the developmental differences on the behavioral level, brain regions typically assigned to the ToM network support distinct processes in adolescents and adults. For instance, medial prefrontal cortex (MPFC) and posterior cingulate cortex (PCC) activity are less related to estimated predictions and more strongly correlated with estimated prediction errors in adolescents compared to adults (see Figure 1 C). 

Conclusions:  

We show that adolescence is marked by substantial developmental changes in social learning. We are currently investigating differences between the developmental trajectories of TD and ASD adolescents. This could help identify behavioral and biological markers of treatment change, which could help predict behavioral treatment outcomes for adolescents with ASD.