A research team from Yonsei University has developed an AI model that screens for attention-deficit/hyperactivity disorder (ADHD) using retinal fundus photographs, achieving an impressive diagnostic accuracy of 96.9 percent during internal testing.

This announcement was made on April 21, 2025, following the analysis of 1,108 retinal fundus images through four learning algorithm models and the innovative AutoMorph Pipeline technology.

The AI model represents a significant advancement in medical diagnostics, providing a more accurate method for diagnosing ADHD and offering insights into neurological and vascular health.

ADHD affects approximately 5-8% of school-aged children and is characterized by attention deficits, impulsivity, and hyperactivity; traditional diagnosis methods often rely on subjective interviews and questionnaires.

These traditional diagnostic approaches can introduce subjectivity and inconsistencies, often leading to delays in diagnosis and treatment.

The research team was led by Professors Chun Geun-ah and Choi Hang-nyung from Severance Hospital, in collaboration with Professor Park Yoo-rang from Yonsei University.

Key retinal features linked to ADHD were identified through Shapley Additive Explanations (SHAP), revealing increased blood vessel density and decreased arterial width as significant indicators.

Professor Cheon emphasized that retinal fundus imaging could serve as a crucial biomarker for diagnosing ADHD and for assessing visual attention deficits.

Current ADHD diagnoses involve lengthy assessments, but the researchers propose that retinal imaging offers a quicker, noninvasive screening method that takes less than five minutes.

The AI model also demonstrated an 87.3% accuracy in predicting impairments in visual selective attention, which is closely associated with executive function in ADHD patients.

Overall, the model exhibited over 91 percent sensitivity and specificity in detecting ADHD, showcasing its potential as a reliable diagnostic tool.

The AI's prediction performance was quantified using the Area Under the Receiver Operating Characteristic Curve (AUROC), achieving a value of 0.969, indicating its high effectiveness.