Machine Learning for ADHD Diagnosis: Feature Selection from Parent Reports, Self-Reports and Neuropsychological Measures

TL;DR

This study shows that machine learning can help diagnose ADHD by identifying key predictors like social problems and reaction time measures, improving clinical decision-making.

Contribution

The study introduces a novel approach using interpretable machine learning to identify robust predictors of ADHD beyond core symptoms.

Findings

Parent-rated social problems, executive dysfunction, and self-regulation were top predictors of ADHD across models.

Ex-Gaussian reaction-time parameters outperformed traditional continuous performance task indices in predicting ADHD.

Abstract

What are the main findings?
Social problems, executive dysfunction, and self-regulation were top predictors of ADHD beyond core symptoms across machine learning models.Ex-Gaussian reaction-time parameters outperformed traditional indices of the continuous performance task.

Social problems, executive dysfunction, and self-regulation were top predictors of ADHD beyond core symptoms across machine learning models.

Ex-Gaussian reaction-time parameters outperformed traditional indices of the continuous performance task.

What is the implication of the main finding?
Prioritizing these key predictors can streamline ADHD assessment and support earlier referral and intervention.Interpretable machine learning models can support clinical decision-making by highlighting the most informative features.

Prioritizing these key predictors can streamline ADHD assessment and support earlier referral and intervention.

Interpretable machine learning models can support clinical decision-making by highlighting the most informative features.

Background: Attention-deficit/hyperactivity disorder (ADHD) is a heterogeneous neurodevelopmental condition that currently relies on subjective clinical judgment for diagnosis, emphasizing the need for objective, clinically applicable tools. Methods: We applied machine learning techniques to parent reports, self-reports, and performance-based measures in a sample of 255 Taiwanese children and adolescents (108 ADHD and 147 controls; mean age = 11.85 years). Models were trained under a nested cross-validation framework to avoid performance overestimation. Results: Most models achieved high classification accuracy (AUCs ≈ 0.886–0.906), while convergent feature importance across models highlighted parent-rated social problems, executive dysfunction, and self-regulation traits as robust predictors. Additionally, ex-Gaussian parameters derived from reaction time distributions on the Continuous Performance Test (CPT) proved more informative than raw scores. Conclusions: These findings support the utility of integrating multi-informant ratings and task-based measures in interpretable ML models to enhance ADHD diagnosis in clinical practice.

Figures