Information-Theoretic Approaches to Optimizing Early Detection of ASD in Toddlers Based on Preferential Attention to Audiovisual Synchrony

Background: Children with ASD exhibit atypical patterns of visual attention to the social world, responding differently to physical and social contingencies relative to non-autistic peers. In previous studies examining preferential attention to audiovisual synchrony, we showed that ASD infants are relatively insensitive to social contingencies afforded by talking faces, focusing instead on physical contingencies between light and sound. By manipulating audiovisual stimuli comprising faces and shapes synchronized with speech and tones, we found that TD controls exhibited a preference for synchronous faces and speech, lacking in ASD participants, even though groups did not differ in baseline sensitivity to audiovisual synchrony. In those studies, significant differences were found based on simple measures of visual attention derived from mean relative fixation durations. Although these measures are traditional in eye-tracking research, inspection of our data clearly revealed complex statistical patterns that were predictive of autism, but were not captured by simple tests of differences in mean fixation on any one region of interest alone. Accordingly, we applied techniques from information theory to quantify differences between full probability distributions of eye-tracking trajectories across groups.

Objectives: Our goal was to test whether information-theoretic measures of differences in visual attention between TD and ASD toddlers outperform traditional statistics, to determine whether current approaches to detection of autism may be significantly under-exploiting the discriminative power of behavioral probes.

Methods: Toddlers with autism (N=34) and typically developing controls (N=20) participated in a simple preferential-looking paradigm based on split-screen presentation of video stimuli (faces and shapes) paired with audio stimuli (speech and tones). Using different combinations of video and audio stimuli, and manipulating audiovisual synchrony between the two, we tested for differences in attention to social and physical contingencies. Eye-tracking was used to quantify response. Using machine learning techniques, we derived optimal measures of overall attention and attention to social target across all stimulus combinations, focusing on responses to speech and non-speech. We used a permutation test to determine significant differences between the entire joint distribution of our measures across groups, based on the information divergence calculated from kernel density estimates. We contrasted these results with traditional measures calculated from relative percentage of total fixation time on the social target, testing for differences in means using ANOVA. Receiver operating characteristics were further used to quantify classification performance.

Results: We found highly significant differences in the full probability distributions (P<0.00001), which were not reflected in tests of differences in centroids alone (P<0.01). Sensitivity and specificity improved from 78.5%/83.3% for single mean fixation measures to 79.5%/97.4% using our information-theoretic approach.

Conclusions: Significant differences in visual scanning exist between ASD and TD children that cannot be fully quantified without characterizing the mutual information between entire probability distributions of responses. Current approaches to screening of children at risk of autism based on group means significantly underexploit information present in behavioral measures of the syndrome.