Item-Level Prediction of ADHD-Related Impairment Using Machine Learning and SEM Across Informants

• In Event: Effect Estimation for Causality and Prediction: Multilevel and Single-case Designs, Machine Learning, and Distributional Assumptions

Abstract

This study identified which ADHD symptom items predict functional impairment across six domains using parent and teacher ratings. Structural equation modeling (SEM) and machine learning (ML; XGBoost, random forest, neural networks) were used to model relationships between 18 DSM-based symptoms and six domain-specific impairments. SEM showed inattention predicted academic, homework and self-esteem impairments, while hyperactivity/impulsivity predicted peer and behavioral problems. XGBoost consistently outperformed other models in F1 and AUPRC metrics. SHAP (SHapley Additive exPlanations) highlighted inattention as most predictive of academic and homework difficulties, with hyperactivity/impulsivity more salient in teacher-rated peer relationship and behavioral outcomes. SHAP heatmaps revealed informant-specific ranking patterns. Findings support combining SEM and interpretable ML for targeted, symptom-level ADHD assessment and intervention.

Author

• Xuefei Zou, University of Pennsylvania