Stacked generalization reveals optimal model complexity for regional EEG-based alcoholism classification

The diagnosis of Alcohol Use Disorder (AUD) currently relies on subjective clinical assessments. While EEG-based machine learning offers an objective alternative, the relationship between data complexity and optimal model architecture remains an open question. This study introduces a novel, systematically stacked ensemble learning pipeline using the UCI-KDD dataset. We propose a data-driven Top-N ranking strategy where twelve diverse base classifiers are ranked by a composite performance score and incrementally added to a stacked meta -model. Our results demonstrate that the full 61-channel EEG dataset achieves a peak accuracy of 96.28% using a Top-9 ensemble. Conversely, lower-dimensional regional data (Prefrontal and Frontal) are optimally classified using a simpler Top-3 ensemble, achieving 90.18% accuracy. These findings reveal a fundamental correlation: as data dimensionality decreases, the optimal ensemble complexity also reduces, providing a preliminary methodological foundation for evaluating the feasibility of efficient, regional EEG-based screening strategies for AUD.