Artificial intelligence in ADHD diagnosis: a comprehensive review of machine learning applications, clinical validation challenges, and implementation barriers in precision medicine

Le résumé fourni par la source

Abstract Background Attention Deficit Hyperactivity Disorder (ADHD) affects approximately 7.2% of the global population, representing one of the most prevalent neurodevelopmental disorders. While current clinical diagnostic approaches, including validated instruments such as the SNAP-IV, Conners-4, and ADHD-RS, enable trained clinicians to reach reliable diagnoses in most cases, certain clinical scenarios remain challenging, including borderline presentations, complex comorbidity patterns, and settings with limited access to specialist clinicians. The absence of objective biomarkers meeting established clinical criteria has motivated exploration of computational approaches as potential augmentative diagnostic tools. Methods This comprehensive narrative review analyzes machine learning (ML) and deep learning (DL) applications across the ADHD clinical pathway, encompassing objective diagnosis, subtype differentiation, prediction of treatment response, digital therapeutic interventions, and long-term outcome prognostication. PubMed/MEDLINE, Scopus, Web of Science, IEEE Xplore, and Google Scholar were searched from inception through September 2025 across multiple modalities including electroencephalography (EEG), neuroimaging, wearable technology, virtual reality, eye-tracking, genomics, digital therapeutics, and multimodal integration approaches. Diagnostic performance, clinical validation, model interpretability, class imbalance considerations, and translational potential were evaluated while identifying critical gaps requiring additional research. Results EEG-based classification demonstrates the most consistent high performance, with recent studies achieving accuracies exceeding 99% using advanced CNN and transformer architectures; however, these results are typically obtained from small, single-site datasets and should be interpreted with caution. Wearable technology shows significant promise with 89% accuracy in large-scale validation ( n = 450; AUC = 0.95), while virtual reality assessments achieve clinically relevant performance (AUC = 0.893). Eye-tracking technology achieves exceptional diagnostic performance (AUC = 0.965, accuracy = 90.8%), and genomic approaches utilizing advanced polygenic risk scores achieve meaningful performance (AUC = 0.72). Multimodal integration approaches demonstrate superior diagnostic accuracy compared to single-modality assessments. Importantly, well-validated studies employing rigorous multi-site validation typically report realistic performance in the range of 70–85%, and metrics such as ROC-AUC, sensitivity, specificity, and positive predictive value (PPV) provide more clinically meaningful evaluations than overall accuracy alone, particularly given the class imbalance inherent in ADHD prevalence (5–10%). Conclusions While ML/DL technologies show considerable promise for augmenting objective ADHD assessment, including diagnosis, treatment response prediction, and long-term outcome prognostication, realistic clinical performance typically ranges from 70 to 85% for well-validated studies. The field requires addressing critical challenges including methodological rigor, clinical validation standards, appropriate handling of class imbalance, gold-standard diagnostic validation, algorithmic interpretability, privacy-preserving techniques, and regulatory frameworks to achieve successful clinical translation. Clinical trial number Not applicable

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.