Machine Learning Predictive Model for ADHD Diagnosis: Presentations and Comorbidities

Background: Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental condition with heterogeneity and clinical variability. It has three diagnostic presentations (inattentive, hyperactive/impulsive, combined). It is common for it to occur concomitantly with other psychopathologies (neurodevelopmental disorder or behavioral disorder), which increases the complexity of its detection. Methods: This work developed and evaluated a multiclass machine learning (ML) model to identify ADHD and comorbidities in a pediatric population. The research approach was quantitative and cross-sectional, with a non-experimental design and non-probability sampling. The sample has 892 children aged [6-12] years from Medellín, Colombia (ADHD=780, typically developing controls=112). Cognitive and behavioral assessments were used as predictor variables. Python 3.12.13 was used in the programming algorithms; a computational pipeline was implemented with preprocessing techniques, class balancing using Synthetic Minority Over-sampling Technique (SMOTE) and stratified cross-validation. Results: Multiple classification algorithms were evaluated, Random Forest (RF) was identified as the optimal model (accuracy=.9778, precision=.9798, recall=.9778, F1-score=.9777; weighted averages for multiclass metrics). The results indicate an adequate capacity of the model to differentiate complex clinical profiles with a performance superior to that of traditional binary classification methods. Conclusions: Use of accessible and clinically interpretable psychometric variables enhances the applicability of the tool in real-world healthcare settings with limited access to resources and specialized professionals. ML is a promising complementary tool for ADHD detection. External validation is needed to confirm its generalizability across different populations and clinical contexts.