Decoding Self-Regulation in Substance Use Disorders: Machine Learning and ORETA Neurofeedback at Precuneus

Substance use disorders (SUD) remain among the most treatment-resistant conditions, particularly in incarcerated populations. Despite decades of neuropsychological models, few approaches capture the dynamic, network-level self-regulation deficits observed in SUD. In this study, 63 incarcerated individuals completed 20 sessions of LORETA neurofeedback targeting alpha (8–13 Hz) at the left precuneus. Pre–post EEG current source density (CSD) and Personality Assessment Inventory (PAI) scores were analyzed. A Random Forest classifier was trained on spectral CSD features and behavioral deltas, with Shapley Additive Explanation (SHAP) values used to interpret model contributions. Results showed significant alpha increases at the trained ROI, accompanied by posterior-to-frontal energy redistribution and functional asymmetry. Notably, both alpha synchronization and desynchronization predicted behavioral improvement, with an overall PAI effect size of d = 0.85. Regression and PCA confirmed directional reorganization and subgroup differentiation. Cohen’s d analysis revealed frequency-specific effects in alpha and low beta bands. Machine learning (ML) revealed that changes in posterior and frontal regions of interest (ROI) were differentially predictive of treatment response, and Reliable Change Index (RCI) identified responders with physiological and behavioral concordance. These findings challenge static trait-based models of addiction and suggest that rhythmic neurofeedback and ML interpretation can uncover emergent regulatory subtypes. This work proposes a shift from amplitude-based training to network-informed modulation as a foundation for scalable, individualized SUD interventions.