The rising prevalence of autism spectrum disorder (ASD) is placing strain on clinical resources. Gold-standard diagnostic instruments such as ADOS-2 require trained clinicians, are costly, and can involve long wait times, delaying diagnosis and early intervention. Digital biomarkers like eye-tracking have been proposed as scalable alternatives or adjuncts to traditional assessment, but prior solutions have faced barriers to community deployment due to specialized hardware costs and operational constraints.
WISE-Screen is presented as a smartphone-based, real-time architecture designed to perform autonomous ASD screening and multidimensional phenotypic profiling using high-fidelity eye-gaze data captured on standard phones. The authors evaluated the conceptual feasibility of this approach across a developmentally diverse age range.
The WISE-Screen framework incorporates two complementary machine learning pipelines that operate on smartphone-captured gaze traces.
The Scanpath-based (SP) pipeline derives features from participant scanpaths across 34 stimuli. It uses saliency maps and engineered scanpath characteristics to estimate the probability of ASD-typical gaze patterns.
The Domain-task-based (DT) pipeline evaluates participant responses across 17 specialized tasks mapped to four phenotypic domains: social, emotional, sensory, and executive. The DT pipeline aggregates task-level information to profile multidimensional phenotypes relevant to ASD.
Both pipelines were implemented to run in real time within the smartphone platform, enabling autonomous capture and preliminary analysis without the need for external eye-tracking hardware.
Model development and evaluation used data from 35 participants aged 2.5 to 17 years: 16 with ASD and 19 without ASD. Diagnostic status for participants was confirmed using ADOS-2. Models were assessed using leave-one-out cross-validation across the full sample.
A baseline demographic model was included for comparison. The study examined individual pipeline performance (SP and DT), their integration, and models adjusted for age and sex.
Performance was reported using receiver operating characteristic area under the curve (ROC-AUC) with 95% confidence intervals.
When accounting for age and sex (residualized models), the adjusted ROC-AUC was 0.74 (95% CI: 0.57–0.92).
These results indicate that both the SP and DT approaches individually improve discrimination relative to demographic predictors, and that an integrated approach achieved the highest ROC-AUC in this sample.
Within the domain-task framework, the sensory, social and emotional domains demonstrated the strongest associations with ASD status in the evaluated dataset. The authors report that age- and sex-residualized models retained predictive signal but with reduced adjusted ROC-AUC compared with unadjusted models, suggesting partial influence of developmental and demographic factors on the measured gaze-derived features.
WISE-Screen is proposed as a scalable, automated adjunct to traditional ASD screening and diagnostic workflows. By leveraging smartphone cameras and on-device analytics, the framework aims to reduce reliance on specialized eye-tracking hardware and broaden access to digital phenotyping in community settings.
The authors emphasize that the current results demonstrate conceptual feasibility rather than definitive clinical utility. The study cohort was small (35 participants) and the authors state that further evaluation in larger, more diverse cohorts is warranted before deployment as a screening tool. Details such as operational performance in uncontrolled environments, longitudinal predictive value, and real-world integration pathways were not reported in this preprint.
The study received ethical approval from the Hong Kong Human Research Independent Review Board (HK-HR-IRB). Participant diagnostic confirmation used ADOS-2, and participant consent and oversight statements are reported.
Funding was provided by WellMind BioMed Technology Holdings Limited, which also holds the patent for the WISE-Screen App referenced in the study. Several authors were employees of WellMind BioMed Technology Holdings Limited, and other authors served as unpaid consultants; these relationships are declared as competing interests.
Participant-level data that support the findings are available upon request from the corresponding authors (Hon-Cheong So and Stephen Kwok-Wing Tsui) but are not publicly available because of privacy protections. The preprint notes that supplementary files and links are provided, and that reuse is restricted by the copyright holder.
In this development-stage study, the WISE-Screen smartphone eye-tracking framework produced promising discrimination between ADOS-2–confirmed ASD and non-ASD participants using scanpath and domain-task features, with an integrated ROC-AUC up to 0.91 in a 35-participant sample. The approach aims to provide scalable, automated screening and multidimensional phenotyping using smartphone-captured gaze, but further validation in larger cohorts and real-world settings is required before clinical implementation.