Severance Professor Chun Geuna's Team Develops AI to Screen for ADHD Using Retinal Fundus Images

Photo by Severance Hospital

A research team led by Professor Chun Geuna and Professor Choi Hangnyeong from the Department of Child and Adolescent Psychiatry at Severance Hospital, together with Professor Park Yoorang from the Department of Biomedical Systems Informatics at Yonsei University College of Medicine, announced on April 21 that the accuracy of an AI system designed to screen for ADHD (Attention Deficit Hyperactivity Disorder) using retinal fundus images reached 96.9%.

The results of this study were published in the international journal 'npj Digital Medicine' (IF 12.4).

Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder affecting 5-8% of school-aged children. The main symptoms are inattention, impulsivity, and hyperactivity. Delays in diagnosis and treatment can affect academic performance, social relationships, and emotional development.

Since ADHD diagnosis relies on interviews and questionnaire assessments, there is a high possibility of examiner subjectivity influencing the results. The distinction between normal behavior and symptoms is often unclear, and inconsistent diagnoses can occur, making rapid diagnosis challenging.

The research team developed an AI system capable of objectively and rapidly screening for ADHD using retinal fundus images.

The development of this AI involved 1,108 retinal fundus images, four different learning algorithm models, and the AutoMorph Pipeline technology. The AutoMorph Pipeline is a research tool that morphologically analyzes retinal blood vessels.

The predictive performance was outstanding, with the AUROC (Area Under the Receiver Operating Characteristic Curve) value of the graph comparing sensitivity (true positive rate) and specificity (false positive rate) reaching as high as 0.969 (accuracy 96.9%). An AUROC value closer to 1 indicates better performance.

Through SHAP (Shapley Additive Explanations) analysis, which explains the AI model's predictions, the team identified major retinal features associated with ADHD. These included increased vascular density, decreased arterial vessel width, and structural changes in the optic disc.

The team also measured the accuracy with which the AI could predict the degree of impairment in visual selective attention by analyzing retinal fundus images of ADHD patients. Visual selective attention refers to the ability to focus on specific areas, and this ability is typically reduced in ADHD patients. The AI achieved an accuracy of 87.3% in this prediction.

Professor Chun Geuna stated, "This study not only demonstrates the potential for retinal fundus images to serve as an important biomarker for ADHD diagnosis, but also confirms that it is possible to predict executive function deficits such as visual selective attention. Fundus examination is extremely simple, taking less than five minutes, and its rapid assessment may also be useful for monitoring the effects of ADHD medications."

This research was supported by the National Information Society Agency of Korea.

This content was produced with the assistance of AI translation services.