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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Attention-deficit hyperactivity disorder (ADHD) is a complex neurodevelopmental disorder requiring personalized treatment due to its heterogeneity.
• Existing predictive models for ADHD treatment response often lack transparency, hindering clinical trust and adoption.
• Early and effective intervention is crucial for mitigating the long-term effects of untreated ADHD.

Purpose of the Study:

• To introduce a novel, explainable machine learning framework for predicting neurofeedback treatment response in individuals with ADHD.
• To enhance personalized ADHD intervention by providing transparent and clinically actionable insights into treatment prediction.
• To identify key predictive features for neurofeedback treatment response using advanced feature selection and explainability techniques.

Main Methods:

• Utilized a dataset of 72 ADHD patients from the TDBRAIN database, including 78 features (demographic, behavioral, NEO-FFI).
• Employed a hierarchical feature selection approach: initial statistical filtering followed by Sequential Forward Selection (SFS) with four reduction methods.
• Developed and evaluated predictive models using five classifiers (Random Forest, SVM, Logistic Regression, ANN, Adaptive Boosting), with SHAP values for interpretability.

Main Results:

• A hierarchical feature selection approach, culminating in SFS, identified seven optimal features that significantly improved the Random Forest model's accuracy to 88.3% ± 6.8%.
• Key predictive features included specific NEO-FFI questions and education level, with five of the seven SFS features identified as highly significant by SHAP values.
• The explainable AI framework provided both global feature importance and local explanations, demonstrating model consistency and highlighting critical predictive factors.

Conclusions:

• The developed transparent machine learning framework achieves high prediction performance for ADHD neurofeedback treatment response, surpassing previous studies.
• The explainable nature of the model fosters trust and supports personalized, data-driven medical decisions for ADHD intervention.
• This approach moves beyond 'black-box' predictions, offering clinically actionable insights for tailoring neurofeedback therapies to individual ADHD patients.