Automated AI based identification of autism spectrum disorder from home videos

Abstract

Autism spectrum disorder (ASD) is a prevalent childhood-onset neurodevelopmental condition. Early diagnosis remains challenging by the time, cost, and expertise required for traditional assessments, creating barriers to timely identification. We developed an AI-based screening system leveraging home-recorded videos to improve early ASD detection. Three task-based video protocols under 1 min each—name-response, imitation, and ball-playing—were developed, and home videos following these protocols were collected from 510 children (253 ASD, 257 typically developing), aged 18–48 months, across 9 hospitals in South Korea. Task-specific features were extracted using deep learning models and combined with demographic data through machine learning classifiers. The ensemble model achieved an area under the receiver operating characteristic curve of 0.83 and an accuracy of 0.75. This fully automated approach, based on short home-video protocols that elicit children’s natural behaviors, complements clinical evaluation and may aid in prioritizing referrals and enabling earlier intervention in resource-limited settings.