Gaze response to dyadic bids at 2 years related to outcomes at 3 years in autism spectrum disorders: a subtyping analysis

Abstract

Variability in attention towards direct gaze and child-directed speech may contribute to heterogeneity of clinical presentation in toddlers with autism spectrum disorders (ASD). To evaluate this hypothesis, we clustered sixty-five 20-month-old toddlers with ASD based on their visual responses to dyadic cues for engagement, identifying three subgroups. Subsequently, we compared social, language, and adaptive functioning of these subgroups at 3 years of age. The cluster displaying limited attention to social scenes in general exhibited poor outcome at 3 years; the cluster displaying good attention to the scene and to the speaker's mouth was verbal and high functioning at 3 years. Analysis of visual responses to dyadic cues may provide a clinically meaningful approach to identifying early predictors of outcome.

Figures

Figure 1

(a) Single frame from the video stimulus. (b) Regions of interest (ROIs) used in analysis for the frame in Figure 1a. Labeled regions are Eyes (E), Mouth (M), Body (B), Hands Area (H), Toys (T), and Background (BG). The scene subtended 27 × 21 degrees of visual angle, the Face 3.9 × 5.6, Mouth 3.5 × 2.0, and the individual Toys 5.8 × 6.4.

Figure 1

(a) Single frame from the video stimulus. (b) Regions of interest (ROIs) used in analysis for the frame in Figure 1a. Labeled regions are Eyes (E), Mouth (M), Body (B), Hands Area (H), Toys (T), and Background (BG). The scene subtended 27 × 21 degrees of visual angle, the Face 3.9 × 5.6, Mouth 3.5 × 2.0, and the individual Toys 5.8 × 6.4.

Figure 2

(a) Dendrogram reflecting the cluster tree based on three performance measures: proportion of time spent attending to the scene (%Scene), Eyes (%Eyes) and Eye-to-Mouth ratio. (b) Height differences in the cluster dendogram for each additional cluster added. The largest difference is made when moving from two to three clusters.

Figure 2

(a) Dendrogram reflecting the cluster tree based on three performance measures: proportion of time spent attending to the scene (%Scene), Eyes (%Eyes) and Eye-to-Mouth ratio. (b) Height differences in the cluster dendogram for each additional cluster added. The largest difference is made when moving from two to three clusters.

Figure 3

(a) Histogram of the optimal number of clusters based on 1,000 bootstrapped samples. 66% of the samples had an optimal cluster number of 3, consistent with our reported findings. (b) The top figure displays a histogram of the values of cluster consistency, as measured by the adjusted Rand index of agreement between our reported cluster partition and each bootstrapped cluster partition. The degree of cluster consistency is significantly larger than that observed from random cluster assignment, shown in the bottom figure.

Figure 3

(a) Histogram of the optimal number of clusters based on 1,000 bootstrapped samples. 66% of the samples had an optimal cluster number of 3, consistent with our reported findings. (b) The top figure displays a histogram of the values of cluster consistency, as measured by the adjusted Rand index of agreement between our reported cluster partition and each bootstrapped cluster partition. The degree of cluster consistency is significantly larger than that observed from random cluster assignment, shown in the bottom figure.

Figure 4

Looking time ratios for each of the indicated regions of interest by cluster group. P-values reported at the top of each plot are from ANOVA F-tests for differences in mean. E/M = eye-to-mouth ratio. Pairwise significances are indicated according to the following legend: *: p<0.05; **: p<0.01; ***: p<0.001.

Figure 5

Results of mixed-effects models predicting each outcome variable using Visit, Cluster, Visit-by-Cluster interaction, and intensity of treatment between visits 1 and 2. Marginal means (+/-1 standard error) after accounting for effects of treatment intensity are plotted for each outcome variable and cluster.

Figure 6

Repeated-measures Cohen's d effect sizes for the magnitude of change in each outcome variable from Visit 1 to Visit 2, performed separately on each identified cluster. All effect sizes are computed over marginal means, controlling for intensity of intervention. Vertical lines indicate 95% confidence intervals for the effect sizes.