Neural Noise- An explainable machine learning approach can predict treatment response for neurofeedback in ADHD 1.
- The method offers predictive biomarkers for neurofeedback; published in Nature (tier-2, clinical ADHD focus) 1. 1
Weekly enrichment (2026-07-20)
- The study (Scientific Reports, 2025) introduces an explainable AI (XAI) framework to predict neurofeedback treatment response in ADHD, aiming for transparent, clinically actionable insights rather than black-box predictions.23
- Data came from 72 ADHD patients (ages 6-68) drawn from the TDBRAIN (two-decades brainclinics archive for insights in neurophysiology) database, using 78 features spanning demographics, behavioral data, and the 60-item NEO Five-Factor Inventory (NEO-FFI) personality questionnaire.23
- Notably, this analysis used only demographic, behavioral, and questionnaire features (not the EEG signals), and 47 of the 72 patients were classified as neurofeedback responders.2
- A two-stage hierarchical feature selection was used: a preliminary statistical screen kept 20 features with AUC > 0.60, then four methods (mutual information, ReliefF, mRMR, and sequential forward selection) refined the set; SMOTE addressed class imbalance.2
- Five classifiers were compared (logistic regression, SVM, random forest, ANN, AdaBoost); random forest performed best, reaching 85.1% (±7.7) accuracy, F-measure 84.4%, and AUC 0.92 with the 20-feature set.2
- With sequential forward selection reducing to just 7 optimal features (six NEO-FFI items plus education), random forest accuracy rose to 88.3% (±6.8).2
- The most discriminative single features were NEO-FFI items neoFFI_q6 (AUC 0.71, p=0.001), neoFFI_q55 (AUC 0.69, p=0.001), and neoFFI_q60 (AUC 0.68, p=0.003), alongside age, education, and sleep.2
- SHAP (SHapley Additive exPlanations) with a TreeExplainer was applied to the random forest for global feature-importance ranking, providing interpretability intended to support personalized, non-pharmacological ADHD intervention decisions.23