Early prediction of childhood ADHD using prenatal and early postnatal behavioural features: evaluation across six machine-learning algorithms

Limited attention has been devoted to developing machine-learning models that use behavioural data for the early prediction of childhood attention-deficit/hyperactivity disorder (ADHD), particularly in the United Kingdom. Therefore, this study evaluated the predictive performance of six machine learning approaches in a cohort of 9,385 children (259 with ADHD, 9,126 controls) from the UK Millennium Cohort Study. After selecting the optimal model, we comprehensively compared the relative contributions of prenatal and postnatal (0–3 years) multi-domain features to its predictive performance. Results indicated that XGBoost achieved the highest performance on the test set (AUC = 0.881), effectively balancing the rates of false positives and false negatives. Specifically, "Conduct problems" is the most significant predictor across all models. Among postnatal features, early childhood cognitive and behavioural development represented the most influential domain, contributing approximately 51.9% SHAP value to the model. Nonetheless, other domain features (e.g. prenatal features) show non-negligible contributions. By establishing robust predictive performance, this research addresses an existing gap in machine learning-based studies of childhood ADHD within the UK context. Furthermore, as the first study to quantitatively evaluate the contribution of multiple behavioural domain features to predictive model performance in ADHD, this work provides valuable insights for future model development.