A MACHINE LEARNING APPROACH TO PREDICT THE THERAPEUTIC EFFICACY OF MOBILE NEUROFEEDBACK IN CHILDREN WITH ADHD

Abstract Background Attention-Deficit/Hyperactivity Disorder (ADHD) is a type of neurodevelopmental disorder characterized by inattention, impulsivity, and hyperactivity, associated with adverse long-term functional outcomes. Neurofeedback training as an add-on to medication therapy has been proposed. Aims & Objectives This study aimed to evaluate the effectiveness of machine learning (ML) models in predicting the therapeutic response to mobile neurofeedback (MNF) in children diagnosed with ADHD. Method: A total of 41 children with ADHD were allocated to one of two groups: ADHD MNF-medication (n=28) and ADHD MNF alone (n=13). Participants already receiving medication for ADHD maintained their medication dosages throughout the MNF treatment period. The MNF program was administered three days a week, twice daily, for 10-20 minutes over a span of 12 weeks. The assessment of the MNF's effectiveness was based on changes in the Clinical Global Impression-Severity scale (CGI-S). We quantified absolute and relative power across 19 channels and analyzed quantitative electroencephalography (qEEG) data within the following frequency ranges: delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (12–25 Hz), high beta (25–30 Hz), and gamma (30–80 Hz). We employed the XGBoost algorithm to determine the classification potential based on quantitative EEG characteristics and achieved optimal classification performance through hyperparameter optimization, utilizing Python- based XGBoost software. Results The highest classification performance obtained through hyperparameter optimization yielded an accuracy rate of 99.7%. To assess the contribution of individual features to the classification, we quantified the impact on classification performance for each group using SHAP (SHapley Additive exPlanations) values and identified the top 10 contributing features. Notably, in the classification based on EEG data obtained with eyes closed, the following features exhibited high importance: F7 (Delta), Fp1 (Delta), F8 (Low Beta), F8 (Beta), F7 (Relative Delta), F4 (Middle Beta), F4 (Beta), Fp1 (Gamma), F7 (Gamma), and O2 (Middle Beta / Theta). Discussion & Conclusion Our machine learning analysis demonstrated a high level of accuracy in predicting the therapeutic response to mobile neurofeedback. Notably, delta and beta waves in the frontal lobe were identified as having the highest impact on classification performance.