A new model called Autism-CLIP uses short home videos to improve early autism detection in toddlers, potentially revolutionising current screening methods with increased accessibility and efficiency.
Researchers have developed Autism-CLIP, a model that uses short home videos to help screen toddlers for autism more quickly and cheaply than traditional methods, according to a Nature Communications study. The system analyses three-minute caregiver-child interaction clips from children aged 15 to 24 months and looks for behavioural patterns linked with autism, including reduced eye contact and weaker social back-and-forth. By combining vision and language modelling with training methods that reduce the need for hand-labelled examples, the researchers say the approach could help widen access to early screening.
In tests across multiple centres, Autism-CLIP produced strong results, with an area under the curve of 0.903 in an internal group of 103 children and 0.873 in an external group of 102. That performance outpaced conventional machine-learning baselines, the study said. The system was designed to work from smartphone video alone, without controlled laboratory settings or manual coding, which could make it easier to use in routine paediatric settings or at home.
The study also points to a broader shift in autism research towards tools that can support, rather than replace, clinical judgement. Current screening often depends on specialist observation, structured assessments and long wait times, which can delay support during a critical period of development. A low-cost digital method, if validated further, could help clinicians prioritise referrals and focus resources on children most likely to need timely evaluation.
The authors said their model uses contrastive learning and large language model-generated supervision to connect visual behaviour with autism-related assessment outcomes. That design lets the system make predictions without the kind of manual annotation usually needed to train medical video models. The researchers said the next step is to confirm how well the tool performs in broader real-world populations and different care settings before it can be considered for routine use.
Disclaimer: This content is for informational purposes only and is not intended to be a substitute for professional medical judgment, advice, diagnosis, or treatment.
