How AI System by Researchers in South Korea Can Help Predict ASD Severity

Researchers at Yonsei University College of Medicine (YUCM) have successfully developed a deep learning-based model capable of diagnosing Autism Spectrum Disorder (ASD) and determining its severity. By focusing on joint attention behaviors, the team aimed to provide an objective and quantitative measure of ASD, addressing the current reliance on subjective clinical assessments. Their findings have the potential to improve diagnostic accuracy and assist in the development of precise screening tools for ASD and other neurological disorders.

Yonsei University, located in South Korea, offers nearly 1,000 courses taught in English and many of them are open to CIEE students in all departments EXCEPT medicine, nursing, law, music, and courses at the remote International Songdo Campus. Many international students choose to study medicine in Korea with the desire to hone more skills and develop careers. However, it is not easy to study in Korea, so you need to have previous academic preparation.

Objective assessment of joint attention

Autism spectrum disorder (ASD) is a neurological and developmental disorder that affects how people interact with others, communicate, learn, and behave. Although autism can be diagnosed at any age, it is described as a “developmental disorder” because symptoms generally appear in the first 2 years of life. Scientists believe there are multiple causes of ASD that act together to change the most common ways people develop.

Joint attention, involving behaviors such as eye contact, head-turning, and eye shifting during social interactions, is a critical behavioral marker for diagnosing ASD. However, the lack of an objective, quantitative measure for joint attention has posed challenges to accurate diagnosis. To address this, Lecturer Ko Chan-young and Professor Park Yu-rang from YUCM’s Department of Biomedical Systems Informatics developed a protocol to objectively assess joint attention behaviors.

Data collection and deep learning model

The researchers recorded videos of joint attention behaviors from 95 children, including 45 with ASD and 50 without ASD symptoms, aged 24-72 months. These videos encompassed three types of joint attention: initiation of joint attention (IJA), low-level response to joint attention (RJAlow), and high-level response to joint attention (RJAhigh). Using this dataset, the team trained a deep learning model to distinguish children with ASD from those without and predict the severity of ASD.

The researchers evaluated the model’s performance using several indicators, including the area under the receiver operating characteristic curve (AUROC), accuracy, precision, and recall. The AUROC measures the diagnostic accuracy of a test, while accuracy represents the percentage of correct predictions in the dataset. Precision assesses the probability of correctly identifying individuals with a positive test result, while recall determines the probability of correctly identifying individuals with the condition.

Impressive results and severity assessment

The deep learning model achieved exceptional performance across all three types of joint attention behaviors. When analyzing IJA videos as input, the model demonstrated an AUROC of 99.6%, accuracy of 97.6%, precision of 95.5%, and recall of 99.2%. The model’s performance remained consistent when analyzing the other two types of joint attention behaviors. 

Regarding severity assessment, which categorizes the severity of ASD into no symptoms, mild-moderate, and severe, the model achieved an AUROC of 90.3%, accuracy of 84.8%, precision of 76.2%, and recall of 84.8% when analyzing IJA videos.

Significance and future implications

The results suggest that IJA is a more precise classification that reflects children’s desire and intentionality for social interaction compared to other joint attention behaviors. Professor Park highlighted the significance of their research, stating that the developed model enables the objective digitization of children’s behavioral indicators to identify ASD and assess symptom severity. The findings have implications for the development of screening tools and precise diagnostic methods for ASD and other neurological disorders with behavioral complexities.

Yonsei University researchers have made significant progress in the analysis of ASD severity by developing a deep learning-based model. Their objective assessment of joint attention behaviors offers a more accurate and quantitative approach to diagnosing ASD and predicting its severity. This breakthrough has the potential to aid research and clinical evaluation, assisting in the development of screening tools and diagnostic methods for ASD and other neurological disorders with similar behavioral challenges. The findings published in the JAMA Network Open contribute to advancing the understanding and management of ASD.

Note: The featured image is by Iris, a child who has autism and expressed herself through painting. Her parents sell her artwork to raise money for her private therapy and artwork. Grace also has Facebook page handled by parents called “@IrisGracePainting”. She has sold several artworks in the UK and around the world.