Accuracy of a 2-minute eye-tracking assessment to differentiate young children with and without autism.

Background: Eye-tracking could expedite autism identification/diagnosis through standardisation and objectivity. We tested whether autism assessment, with Classification Algorithm derived from gaze fixation durations, would have good accuracy (area under the curve [AUC] ≥ 0.80) to differentiate 2-4-year-old autistic from non-autistic children.

Methods: Community sampling (March 2019-March 2021) of 2:00–4:11 year-olds included children recruited into a diagnosed Autism Group (‘cases’) and Non-Autism ‘Control’ Group (with likely undiagnosed autism minimised). We recruited well beyond minimum necessary sample size to ensure within-group heterogeneity and allow exploratory subgroup analysis. Alongside eye-tracking attempted with all recruited participants, we collected parent-report measures for all children, and clinical/behavioural measures with autistic children.

Results: 102 autistic (81.4% male; = 44mths;  = 8.8) and 101 non-autistic children (57.4% male;  = 40;  = 10.5) were recruited and eligible; the former slightly older, proportionately more male, and reflecting greater socio-demographic diversity. autism assessment was completed with 101 non-autistic children ( = 1 returning minimal data), and attempted with 100- and completed with 96 autistic children ( = 2 not attempted following adverse responses to clinical testing;  = 4 attempted but unable to calibrate). The Non-Autism Group returned significantly more overall tracking data. The final Classification Algorithm (range 0-100; threshold score = 28.6)—derived from  = 196 children’s fixation durations to elements of social/non-social scenes, human face presentations, and referential attention trials—had AUC = 0.82 (sensitivity = 0.82, specificity = 0.70). Compared to those correctly classified, autistic children misclassified as ‘controls’ showed greater overall tracking, and less pronounced autism features and developmental disability. Compared to correctly classified non-autistic children, those misclassified as ‘cases’ were older with lower overall tracking.

Limitations: Our groups differed on socio-demographic characteristics and overall tracking (included within the Classification Algorithm). We used the ‘Scene 10A’ stimulus set as provided, without update/modification. Industry employees who developed the final Algorithm were non-blinded to child group, and considered only gaze fixation durations. Community sampling and ‘case-control’ design—comparing diagnosed autistic vs. non-autistic children—could be improved via future referral-based recruitment.

Conclusions: Most children tolerated autism assessment, and our Classification Algorithm properties approached those reported from other use and established clinical assessments. Independent replication is required, and research informing the most suitable clinical application of this technology.

Trial Registration: ACTRN12619000317190

Supplementary Information: The online version contains supplementary material available at 10.1186/s13229-025-00670-4.