Enhanced Performance of the Early Detection of Attention Deficit Hyperactivity Disorder Using Artificial Intelligence

Abstract

Abstract : Attention Deficit Hyperactivity Disorder (ADHD) is a highly prevalent neurodevelopmental condition characterized by persistent patterns of inattention, hyperactivity, and impulsivity. Despite its prevalence, ADHD remains largely underestimated in society and is often mistaken for naughtiness in children's attitudes. This is aggravated by the fact that ADHD diagnosis is still largely dependent on subjective clinical assessments, which often result in delayed diagnosis, misdiagnosis, and untreated symptoms. This study investigates the potential use of various Artificial Intelligence (AI) techniques, including traditional machine learning (ML), deep learning (DL), and ensemble methods, to enhance the early detection of ADHD, especially in children. This is accomplished through a comparative evaluation of different data modalities such as handwriting, drawing trajectories, social media posts, electroencephalography (EEG) signals, and neuroimaging. The experimental framework systematically evaluated each modality with rigorous preprocessing, hyperparameter optimization, and cross-validation On handwriting data, support vector machine (SVM), XGBoost, and a stacking ensemble achieved equivalent top performance (accuracy ≈ 90%, F1-score = 0.9, recall = 1.0), with SVM offering the optimal balance of accuracy and computational efficiency. On drawing-based classification, K-Nearest Neighbors (KNN) delivered the highest discriminative performance, with optimal results on all metrics (accuracy = 100%, F1-score = 1.0), followed by SVM at 95.45%, suggesting high potential for behavioral-task-based screening. For textual data, transformer-based language models significantly outperformed statistical and sequential baselines. MentalRoBERTa achieved 87.22% accuracy (F1-score = 0.873), with ensemble strategies providing slight but consistent improvements. EEG-based classification favored ensemble methods integrating KNN, SVM, and Random Forest, achieving 93.26% accuracy and near-perfect recall (0.97), making them suitable for confirmatory screening in clinical settings. Neuroimaging classification using transfer learning demonstrated that ResNet50V2 achieved 83% accuracy, highlighting the limited effectiveness of MRI-based ADHD detection given the challenges in MRI data collection. The results collectively support a staged diagnostic pathway that begins with handwriting and drawing analysis for low-cost, non-invasive screening, incorporates digital text analysis for longitudinal monitoring, utilizes EEG as an objective confirmatory modality, and lastly reserves neuroimaging for complex mental disorder cases. Emphasis is placed on models with high recall to minimize false negatives and support early intervention. The study also highlights interpretability, computational efficiency, ethics, and future research in multimodal fusion and privacy-preserving learning. Hopefully, this work will contribute to the growing body of research on AI-assisted mental health diagnostics and bring more awareness to society about this widespread yet overlooked problem