Development of an Autism Risk Index Using Remote Eye Tracking to Social Stimuli: A Preliminary Proof-of-Principle Investigation

Background: Deficits in eye gaze are a hallmark characteristic of autism. Previous research has identified gaze abnormalities to a range of social stimuli, with large differences between autism spectrum disorder (ASD) and healthy control populations. Yet, no studies have examined whether eye gaze to social stimuli can sensitively differentiate individuals with ASD versus those with non-ASD diagnoses - the most clinically-relevant comparison.

Objectives: This preliminary investigation used remote eye tracking to identify expected social attention at key time points and spatial locations across distinct stimuli. Using expected social attention patterns, the primary objective was to develop an autism risk index to aide early identification of ASD.

Methods: Participants were recruited after initial evaluation for ASD. Consensus diagnoses were derived by multi-disciplinary evaluation and based on the Autism Diagnostic Observation Schedule and clinical interview. Eye gaze was collected using the SMI Red-m remote eye tracker (120Hz sampling) during viewing of static and dynamic social stimuli: emotive faces, biological point-of-light motion, and scenes eliciting predictive gaze and theory of mind. Regions-of-interest at pre-specified time intervals (temporal ROIs) were identified within each stimulus based on a priori hypotheses regarding eye gaze to the social context. Proportion dwell time was calculated to each temporal ROI. To develop the risk index, we first examined the significance and magnitude of ASD versus non-ASD differences to the temporal ROIs within each stimulus. Next, we aggregated this information across all stimuli using the results of a bootstrapped logistic regression with dwell times to each temporal ROI as predictors and diagnostic status as the dichotomous outcome. Finally, we computed classification accuracy, areas under the curve, sensitivity, and specificity.

Results: The sample included 20 ASD (Mage=4.96, range=1.8-7.9; 75% male; ADOS total score range=6-23; Receptive language SS range=55-150) and 13 non-ASD participants (Mage=4.02, range=2.5-6.7; 69% male). All ASD vs. non-ASD differences were in the expected direction and a substantial proportion (12/38) of temporal ROIs showed significant and large-to-very large differences (Cohen’s d=0.62-1.82). Logistic regression identified 8 temporal ROIs as significant unique predictors of ASD status (X²=31.96, R²=.84, p<.00001, 87.9% classification accuracy). Aggregating these predictors into a single risk index yielded high diagnostic discrimination (AUC=.98; 95% CI=.93-1.00) with good sensitivity (.90) and specificity (.93) for the optimal cut score (Figure 1). The autism risk index was highly correlated with ADOS scores (r=.62, p<.001) and, as hypothesized, modestly, but not significantly, correlated with language ability (r=-.32, p=.078).

Conclusions: These findings represent an initial step toward development of an early objective autism risk marker in a challenging clinical sample with broad symptom and cognitive levels. Strong validity within individual stimuli highlights the importance of pinpointing expected spatially- and temporally-dynamic social attention patterns rather than using total dwell times across an entire stimulus. The results, while bootstrapped, will require replication in a large (N≥80 per group) validation sample that includes a broad range of ASD severity and cognitive ability. If validated, the eye gaze risk index may be a useful screening tool and adjunct to clinical judgment in the diagnostic process.