An explainable machine learning-based approach to predicting treatment response for neurofeedback in ADHD

Attention-deficit hyperactivity disorder (ADHD) is a common neurodevelopmental disorder with serious long-term effects if untreated, emphasizing the need for early treatment given its neurobiological heterogeneity. This study introduces a novel explainable machine learning framework to predict neurofeedback treatment response for personalized ADHD intervention, offering transparent and clinically actionable insights. Seventy-eight features, including demographic, behavioral, and personality questionnaire data from 72 ADHD patients (aged 6-68) from the two-decades brainclinics (TDBRAIN) database, were used. First, a preliminary statistical analysis selected 20 features, comprising NEO five-factor inventory (NEO-FFI) questions, behavioral, and demographic data (age, education, sleep) for further analysis. Then, four feature reduction methods, mutual information, ReliefF, minimal-redundancy-maximal-relevance, and sequential forward selection (SFS), were utilized to select the best features. Five classifiers, random forest (RF), support vector machine, logistic regression, artificial neural network, and adaptive boosting, were used to predict neurofeedback treatment response in individuals with ADHD. Subsequently, shapley additive explanations (SHAP) values via TreeExplainer were crucial for interpretability, providing global feature importance and local explanations on model predictions. The results revealed that a hierarchical feature selection approach involved initial statistical filtering, followed by the SFS method, significantly improved the RF model's discrimination to 88.3 ± 6.8% accuracy with just seven optimal features, including NEO-FFI questions and education. Notably, five of the seven SFS features were among SHAP's top 10 most significant, demonstrating the internal consistency of the model and highlighting the features most critical to the model's prediction. Therefore, this transparent machine learning approach achieves competitively higher prediction performance than previous studies and supports personalized, trustworthy ADHD medical decisions, moving beyond black-box prediction.