Integrating explainable AI with clinical features to enhance ADHD diagnostic understanding.

INTRODUCTION

Attention Deficit Hyperactivity Disorder (ADHD) in adults remains challenging to diagnose accurately, with over- and under-diagnosis common due to reliance on subjective clinical judgement. Although machine learning (ML) tools have shown promise in improving diagnostic accuracy, their limited transparency restricts clinical adoption. Existing research rarely integrates broad clinical, substance-use, and quality-of-life measures into a unified predictive framework, nor does it systematically compare explainable artificial intelligence (XAI) outputs with traditional statistical analyses.

METHODS

We retrospectively analysed 786 anonymised adult assessments (January 2019-December 2024) from a UK specialist mental health service. The dataset included demographics; validated symptom scales (MDQ, GAD-7, PHQ-9, CAARS, DIVA); substance-use screens (AUDIT, DAST); and EQ-5D-3L quality-of-life indices. An XGBoost classifier was trained using a stratified split and evaluated on the held-out test set. Model interpretability was examined using SHapley Additive exPlanations (SHAP). SHAP attributions were triangulated with traditional exploratory analyses, including Pearson correlation matrices and Welch's t-tests, to validate feature relevance and identify interaction effects.

RESULTS

The model achieved 77% accuracy and an AUC-ROC of 0.82. CAARS ADHD Raw scores and DIVA adulthood inattentiveness emerged as the strongest predictors of ADHD diagnosis. SHAP analysis revealed important interaction patterns, including depressive symptom severity (PHQ-9) amplifying the predictive contribution of ADHD symptom scales. Age and gender moderated key feature effects, suggesting demographic variability in symptom expression. Traditional EDA confirmed the statistical significance of these predictors while highlighting complementary linear associations, supporting the robustness of the SHAP-derived explanation profiles.

DISCUSSION: Integrating multimodal clinical features with transparent ML methods provides interpretable, clinically aligned insights into adult ADHD diagnosis. The combined SHAP-EDA approach identifies actionable thresholds, clarifies differential feature contributions, and highlights the importance of comorbidity and demographic context in diagnostic evaluation. These findings support a patient-centred, data-driven framework for improving diagnostic consistency in clinical practice. Future work should focus on multi-site validation and temporal analyses to assess generalisability and stability of feature influences over time.