Explainable Artificial Intelligence Applied to the Characterisation of ADHD in Children and Adolescents Using the WISC-V

Abstract

Attention-Deficit/Hyperactivity Disorder (ADHD) diagnosis relies heavily on subjective behavioural ratings, which are susceptible to informant bias. This study examined whether combining a psychometric cognitive assessment with behavioural rating scales, analysed through Explainable Artificial Intelligence (XAI), could support a more objective and transparent characterisation of ADHD. A clinical dataset of 186 children and adolescents (6–16 years; 125 with ADHD, 61 controls) was assessed using the Wechsler Intelligence Scale for Children—Fifth Edition (WISC-V) and the ADHD Rating Scale-5 (ADHD-RS-5). The ‘Cognitive Gap’ (the discrepancy between the General Ability Index and the Cognitive Proficiency Index) was significantly larger in the ADHD group than in controls (median = 11 vs. 1 points; U = 2155.0, p < 0.001, r = 0.435), a pattern consistent across sex. Following Multivariate Imputation by Chained Equations and Boruta-based feature selection, a Random Forest classifier trained on nine features (five cognitive, four behavioural) achieved 80% accuracy (F1 = 0.86, recall = 0.89), slightly outperforming the model using the full 34-variable set. SHapley Additive exPlanations (SHAP) identified the Cognitive Proficiency Index, parent-rated ADHD-RS-5 total scores, and the Cognitive Gap as the strongest predictors, while Full-Scale IQ contributed negligibly. Cognitive and behavioural measures showed comparable and complementary predictive value. These findings support WISC-V- and XAI-based models as transparent, hypothesis-generating tools that complement, rather than replace, standard clinical diagnosis of ADHD.