Optimization of an Eye-Tracking-Based Categorical Screener for Autism Spectrum Disorders in 18- to 42-Month-Old Children

Background: The overarching goal of this project is to develop performance-based measures that can be used as quantitative diagnostic screeners for autism spectrum disorders (ASD). This research stems from past work in which eye-tracking technology was used to measure spontaneous visual fixations during viewing of naturalistic social situations.  While watching scenes of social interaction, toddlers with autism spent—relative to age- and verbal IQ-matched typically-developing (TD) controls—markedly increased time fixated on mouth, body, and object regions and markedly less time fixated on the eyes.  While these measures showed robust between-group differences, subsequent research identified measures of dynamic visual scanning (moment-by-moment variation in looking patterns) that distinguished the ASD from TD groups with still larger effect sizes.  These results, however, looked at group comparisons, and not at classification of individual children.  In parallel to the current study, we are assessing the validity of visual scanning measures as a categorical screener for ASD in individual children between 18 and 42 months of age (Valente, Ly, Klin, & Jones). The current study aims to optimize the sensitivity and specificity of that screener by constraining the heterogeneity of the training set. 

Objectives: The objective of this study is to test whether constraining the training sample for the initial ASD comparison group will improve the performance of an eye-tracking-based, categorical screener for ASD.    

Methods: Eye-tracking data were collected while 50 toddlers, aged 18 to 42 months, viewed dynamic scenes of other children at play. Standardized clinical assessment measures (ADOS, ADI, cognitive, and language testing) confirmed diagnostic status for ASD and TD children. Rather than using consecutive referrals to identify training and test samples, we selected the ASD training sample (ASD-1, N=50) on the basis of autism severity (focusing on children with lower severity scores). Relative to TD children (N=50), these children provided a training set with which to develop a model of expected differences between ASD and TD visual scanning.  We then tested the remaining ASD children (ASD-2, N=50) as an external validation sample. Receiver operating characteristic (ROC) curves were created to analyze sensitivity and specificity.

Results: Preliminary results indicate that use of the constrained training sample decreased sensitivity but increased specificity of classification. Children in the validation sample were classified with sensitivity of 80% and specificity of 84%.  These results are above recommended benchmarks for Level 1 developmental screeners.  

Conclusions: By constraining the ASD sample to those with lower severity scores, the specificity in classifying individual children increased.  In ongoing analyses, we are now measuring the extent to which additional approaches to optimization (time- and event-based, as well as cohort-based) will improve the classification. These analyses will test the viability of using differences in visual scanning as a categorical screener for autism spectrum disorders.