Explainability-Driven Feature Selection and Ensemble Learning for EEG-Based Alzheimer's Staging

Abstract

Alzheimer's disease is a neurodegenerative disorder characterized by progressive loss of cognitive function. Diagnosis is difficult in the early stages of the disease. Electroencephalography (EEG) data has recently emerged as a low-cost biomarker for determining the stages of the disease. This study comprehensively examined the effectiveness of tree-based machine learning models. Multiclass (Healthy, Mild, Moderate, and Severe) Alzheimer's data obtained from EEG recordings were used. In the proposed method, tree-based machine learning models were trained and compared. The study also applied SHAP analysis within the explainable AI approach to identify the most discriminative EEG features. These features were used to create a reduced feature set, which was then used to build an ensemble learning structure combining the three most successful models. The results showed that the proposed method achieved over 98% accuracy and Macro-F1 performance. The results show that the proposed ensemble model is quite similar to full-featured models despite using approximately half the number of features.These results demonstrate that high-accuracy results can be achieved using fewer features in EEG-based classification of Alzheimer's stages and that the proposed method offers significant advantages in terms of computational cost, interpretability, and clinical applicability. © 2026 IEEE.