Conditional Probabilities of Dynamic Visual Scanning Quantify Altered Pathways of Learning in Toddlers with Autism Spectrum Disorder

Background: Perceived stimulus salience and learning are tightly linked: attending to a particular stimulus at a given moment may result in learning that impacts where one chooses to look in the future. Likewise, not attending to that stimulus may lead to a different pathway of learning, resulting in different future fixations. Despite reports that children with Autism Spectrum Disorders (ASD) attend to social scenes differently than typically-developing (TD) children (Shultz et al., 2011), studies have not examined conditional probabilities of dynamic visual scanning, that is, the probability that viewers will fixate a particular location given viewers’ previous fixation locations. Investigating the conditional probability of fixations may reveal what is learned and what is missed by children with ASD.

Objectives: This study characterizes learning pathways in toddlers with and without ASD by identifying on-screen fixation locations that predict future fixation locations.

Methods: Eye-tracking data were collected from 81 toddlers with ASD (mean age=24.1(1.7) months) and 75 TD toddlers (mean age=24.6(0.9) months) while viewing clips of naturalistic peer interactions. Fixation targets were defined as characters or objects perceived as highly salient (indexed by timepoints when most viewers looked at the same location; Figure 1a-b). Targets were identified separately for TD and ASD groups (‘TD targets’ and ‘ASD targets’). Conditional probability ratios were calculated for all target pairs as the proportion of viewers who fixated on initial target X and later target Y relative to those who fixated on Y but not X (Figure 1c). Conditionally dependent target pairs had conditional probabilities above 1.0, indicating targets were more likely to be fixated on had viewers looked at specific locations, and were less likely to be fixated on had viewers not looked at those earlier locations. Conditionally dependent target pairs were identified for TD viewers fixating on TD and ASD targets, and for ASD viewers fixating on ASD and TD targets.

Results: Conditionally dependent target pairs were identified for TD and ASD viewers, indicating that viewing patterns of both groups are influenced by past viewing experience. Interestingly, TD fixations on TD targets appeared to be more strongly influenced by previous viewing experience compared to ASD fixations on ASD targets: a greater number of conditionally dependent target pairs were identified for TD viewers and the conditional probabilities of identified pairs were higher for TD (mean=5.33(10.40)) than for ASD viewers (mean=3.17(3.15)). When TD toddlers fixated on ASD targets, conditional probabilities were lower than when fixating on TD targets (mean=2.66(2.60)), suggesting that TD fixations are less strongly influenced by fixation locations perceived as salient by ASD viewers. By contrast, conditional probabilities were comparable when ASD viewers fixated on ASD compared with TD targets (mean=3.17(3.15), mean=3.19(3.63), respectively), suggesting ASD fixations are similarly influenced by TD and ASD targets.

Conclusions: In navigating the social world, current understanding depends on past experiences, as information conveyed in past fixation locations may influence future viewing. Calculating conditional probabilities of dynamic visual scanning provides a novel means of quantifying how learning shapes and is shaped by altered patterns of visual engagement in ASD.