Retinal fundus imaging as biomarker for ADHD using machine learning for screening and visual attention stratification

Abstract

Attention-deficit/hyperactivity disorder (ADHD), characterized by diagnostic complexity and symptom heterogeneity, is a prevalent neurodevelopmental disorder. Here, we explored the machine learning (ML) analysis of retinal fundus photographs as a noninvasive biomarker for ADHD screening and stratification of executive function (EF) deficits. From April to October 2022, 323 children and adolescents with ADHD were recruited from two tertiary South Korean hospitals, and the age- and sex-matched individuals with typical development were retrospectively collected. We used the AutoMorph pipeline to extract retinal features and used four types of ML models for ADHD screening and EF subdomain prediction, and we adopted the Shapely additive explanation method. ADHD screening models achieved 95.5%-96.9% AUROC. For EF function stratification, the visual and auditory subdomains showed strong (AUROC > 85%) and poor performances, respectively. Our analysis of retinal fundus photographs demonstrated potential as a noninvasive biomarker for ADHD screening and EF deficit stratification in the visual attention domain.

Fig. 1. AUROC plot of ADHD screening based on retinal features.

AUROC curves of the four machine learning models for distinguishing between ADHD and TD. ADHD, attention-deficit/hyperactivity disorder; TD, typical development; Random Forest, random forest classifier; XGBoost, extreme gradient boosting classifier; Extra Trees, extra tree classifiers. Circle, “X”, and square indicate the AUROC values of eye tracking, electroencephalography, and functional magnetic resonance imaging.

Fig. 2. Feature importance analysis for ADHD classification using SHAP values of XGBoost classifier.

a Beeswarm plot illustrating the directional impact of each feature on model predictions. Positive SHAP values indicate contributions toward ADHD classification, while negative SHAP values indicate contributions toward TD classification. The x-axis represents SHAP value magnitudes, and the color gradient (red to blue) represents the feature values, with red indicating higher values and blue indicating lower values. b Bar plot showing the average magnitude of absolute SHAP values for the top 20 features. Longer bars indicate features with greater overall influence on the model’s predictions. ADHD attention-deficit/hyperactivity disorder, TD typical development, SHAP Shapley additive explanations.

Fig. 3. Performances of distinguishing subdomains of executive function deficits using comprehensive attention tests in ADHD based on retinal features.

Complete results for distinguishing subdomains of executive function deficits using comprehensive attention tests in ADHD based on retinal features. The heatmap depicts the performance based on AUROC values, with red indicating higher performance and blue indicating lower performance. The gray background color indicates no age- and sex-matched undersampling because of the large data size for VSA-RT, ASA-RT, and others or insufficient data for ASA-OE and ASA-CE. AQ values represent the normalized scores for each executive function, with a mean of 100 and a standard deviation of 15. Cutoff values were set in 5-point increments between 70 and 100 for the analysis. These cutoffs were used to divide the ADHD group into two subgroups for executive function stratification. We developed multiple models to distinguish 96 subdomains with 16 AQ values and six incremental cutoff values ranging from 70 to 95 with five points. ADHD attention-deficit/hyperactivity disorder, AUROC area under the receiver operating characteristic curve, AQ attention quotient, VSA visual selective attention, ASA auditory selective attention, OE omission errors, CE commission errors (CE), RT mean reaction time.