An Algorithmic Comparison for Robust ADHD Classification on Imbalanced National Health Survey Data

Machine learning offers a promising avenue for improving the diagnostic classification of neurodevelopmental disorders, yet performance is often hindered by the class imbalance inherent in large-scale health datasets. This study presents an algorithmic comparison of Random Forest (RF) and Extreme Gradient Boosting (XGBoost) to develop a robust classification model for Attention Deficit Hyperactivity Disorder (ADHD) using imbalanced data from the 2023 National Survey of Children’s Health (N=54,749). The primary objective was to evaluate algorithmic performance under various data resampling conditions. A Random Undersampling (RUS) technique was applied at multiple ratios to mitigate class imbalance. Models were systematically evaluated using accuracy, precision, recall, and F1-score. The results demonstrate that while both algorithms achieved high accuracy, XGBoost consistently outperformed RF in balancing precision and recall, particularly on undersampled data. The optimal model was an XGBoost classifier trained on a dataset with a 0.3 undersampling ratio, which achieved a peak F1-score of 84% and an accuracy of 93%. Feature importance analysis confirmed that variables related to special health care needs, medication use, and behavioral diagnoses were the most significant predictors. This research concludes that XGBoost, when combined with moderate undersampling, provides a superior and more robust framework for ADHD classification in imbalanced real-world datasets, highlighting its potential as a tool for clinical decision support.