360. Machine learning-based classification of adult attention-deficit/hyperactivity disorder using cortical structural features
Abstract
Abstract Background Attention-Deficit/Hyperactivity Disorder (ADHD) affects a considerable proportion of adults, yet it remains frequently underdiagnosed. Machine learning approaches using brain imaging features may enhance diagnostic accuracy. Nevertheless, research focusing on adult populations, particularly in Asia, is limited. Aims & Objectives This study aims to evaluate the performance of machine learning algorithms in classifying adult ADHD via analyzing regional cortical morphology. Method A cohort of adults (≥18 years) was recruited from a tertiary medical center in Taiwan. Cortical thickness and surface area metrics were derived from all brain regions using FreeSurfer software. Five distinct machine learning models (K-nearest neighbors, random forest, Gaussian Naive Bayes, logistic regression, and support vector machines) were trained on individual cortical thickness and surface area features, combined with age and sex. The performance of each model was assessed over 100 iterations, using an 80% training and 20% test split, and the mean Area Under the Curve (AUC) was calculated to identify the top-performing model. Results Out of 420 participants, 36.4% had ADHD (mean age ± SD, 26.7 ± 6.5 years; male, 55.3%). The Random Forest model outperformed others for both cortical thickness and surface area analysis. For cortical thickness, the mean AUC was 0.66, with an accuracy of 0.66, a sensitivity of 0.19, and a specificity of 0.74. For surface area, the mean AUC was 0.59, with accuracy of 0.72, sensitivity of 0.06, and specificity of 0.98. Overall, high specificity was observed across models; however, sensitivity was low, highlighting a challenge in accurately identifying true positives. Discussion & Conclusions Cortical structural features exhibited limited ability in distinguishing ADHD in adults. This finding is consistent with previous research and may indicate diminished morphological differences with age. Structural MRI alone proved inadequate for sensitive classification. Future studies may benefit from integrating functional connectivity, neuropsychological features, or medication histories to improve predictive performance.