HKUST unveils AI super agent for autism screening

HKUST unveiled an artificial-intelligence platform named ASD Super Agent that evaluates home-recorded videos to gauge autism risk, suggest personalized interventions and monitor developmental milestones.

How the platform works

Parents upload short clips of everyday activities captured on a smartphone, complete standard questionnaires and provide basic developmental data. The system then processes these inputs to generate a risk assessment and a report useful for professional consultation.

The AI model examines posture-related cues extracted directly on the device, limiting data sent for analysis. It combines video insights with questionnaire responses to produce a single risk score.

In early trials the screening achieved an area under the receiver operating characteristic curve of 0.92, meaning the model correctly ranked an autism-labelled video higher than a non-autistic one about 92% of the time, according to study lead Professor Kani Chen.

Performance and privacy

To safeguard privacy, the platform prioritizes on-device feature extraction, transmitting only the minimal information needed for assessment. This approach reduces exposure of raw video content. By keeping processing on the phone, the tool also limits the requirement for fast internet connections.

The intervention guide follows a “20+1” framework, integrating more than twenty established methods such as Applied Behaviour Analysis, Naturalistic Developmental Behavioural Interventions and TEACCH Structured Teaching. An additional module lets therapists and special-education providers add their own techniques.

Professor Chen emphasized that the system does not indiscriminately merge all methods. Recommendations aim to be coherent, tailored to a child’s goals and practical for parents or teachers to implement.

Broader implications

For ongoing monitoring, the platform builds an evolving profile using text records, periodic video assessments and developmental data supplied by families and clinicians. It can decompose long-term goals into daily tasks and track progress over time. Such continuous tracking helps keep caregivers informed about developmental trends.

Voice interaction supports Cantonese, Putonghua and English, broadening accessibility. Researchers also created a multimodal autism database containing over 1.1 million video, posture and behavioural records to fuel further AI development.

Additional validation studies are required before the tool can be deemed clinically effective for widespread deployment, the university notes.

The project stemmed from community research that highlighted long assessment wait times, high treatment costs and fragmented information about services. Researchers gathered insights during visits to special-education centres, child-development clinics and psychiatric hospitals across Hong Kong, the Greater Bay Area and Beijing, and consulted nearly 500 parents.

Unlike the AI-powered stool test introduced by the Chinese University of Hong Kong, which screens for biological markers, this system focuses on behavioural cues captured in videos, offering a complementary pathway for early detection.

Future real-world testing will examine not only technical performance but also whether families understand the outputs, can follow suggested activities and find the tool useful over time, according to the study lead. Ongoing feedback from users will guide further refinements.

The development team plans to refine the user experience, enhance system performance and expand clinical validation trials.

Institutions could employ the platform to prepare and review intervention plans and analyse video submissions, while leaving final therapeutic decisions to qualified professionals.