Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Early diagnosis of Autism Spectrum Disorder (ASD) is crucial for improving long-term outcomes, yet traditional diagnostic methods are often resource-intensive and can lead to long waiting periods. A recent study has introduced a novel framework that could help address this challenge: a two‑stage, multimodal artificial intelligence system designed to screen for ASD in toddlers as young as eighteen months.
The research, led by a consortium from Seoul National University and Yonsei University in Korea, was the first outcome from a large‑scale national project commissioned by the country’s Ministry of Health and Welfare. The team developed a multimodal AI framework meaning the system analyzes and combines different types of data to form a more complete assessment.
The framework was designed to be accessible. Using a mobile application, researchers collected data from two primary sources during natural parent‑child interactions. The first source was voice data, capturing the audio of the child playing with their parent. The second source was the text data from standard screening questionnaires, including the M‑CHAT, SCQ‑L, and SRS, which are commonly used to gather a parent’s observations of their child’s behavior.
The AI system then analyzes all this information in two distinct stages. In Stage One, the framework works to differentiate between children who are typically developing and those who are either at high risk for ASD or already diagnosed. In Stage Two, the system aims to make a more refined distinction, separating the high‑risk group from those children who have a confirmed diagnosis of ASD.
The study involved a large sample of 1,242 children between the ages of eighteen and forty‑eight months. The results showed strong performance for this approach. In Stage One, the model achieved an AUROC of 0.942, a statistical measure of its ability to distinguish between two groups. In Stage Two, distinguishing high‑risk from diagnosed children, the framework achieved an AUROC of 0.914 and a classification accuracy of 85.2 percent.
Most importantly, the AI’s risk predictions were then compared to the gold‑standard clinical assessment tool for ASD, the ADOS‑2. The model’s predictions strongly agreed with the ADOS‑2 results, showing a highly significant correlation.
This research demonstrates the potential for a scalable, accurate approach to early ASD screening that could be delivered through a mobile platform. By integrating objective voice analysis with information from validated parent‑reported questionnaires, the AI provides a method to help stratify risk in a large population, potentially enabling earlier access to intervention services. The framework is intended to aid in detection and risk stratification, not to replace a comprehensive clinical diagnosis. However, it represents a promising step forward in using digital health innovations to address the critical need for timely and accessible developmental assessments.
