Development and multi-center validation of a school-home integrated machine learning model for early screening of...

Attention-deficit/hyperactivity disorder (ADHD) is frequently under-identified in community settings, delaying necessary intervention. This study aimed to develop and validate a multidimensional Early Screening Prediction Model (ESPM-ADHD) for children aged 6 to 18 years, integrating school-home collaborative data.

A large-scale cross-sectional survey was conducted from January to December 2023 across three distinct regions in China. The dataset was divided into a development cohort (Chongqing, n = 15,085) and two independent external validation cohorts (Shaanxi, n = 7435; Yunnan, n = 4206). Machine learning algorithms, including Random Forest, XGBoost, LightGBM, and Logistic Regression, were trained using 15 selected features. The best-performing model was deployed as a web-based tool.

A total of 26,726 valid questionnaires were analyzed, with an overall ADHD risk prevalence of roughly 2%. Among the tested algorithms, the Random Forest model demonstrated the most robust performance. In external validation, the model maintained high discriminatory power across diverse populations. Specifically, it achieved an AUPRC of 0.298 and an AUROC of 0.929 in the Shaanxi cohort, and an AUPRC of 0.253 and an AUROC of 0.907 in the Yunnan cohort. SHAP analysis identified antisocial behavior, academic performance (native language and math), and daytime sleepiness as the top predictors.

The ESPM-ADHD, constructed using the Random Forest algorithm, effectively identifies children at risk for ADHD by leveraging easily assessable demographic, academic, and somatic indicators. The developed online calculator provides a practical, objective triage tool for clinicians and guardians to facilitate timely referral and diagnosis.