Automated detection of Alzheimer's disease from EEG signals using VMD driven multi domain features and rich and poor optimization enhanced classification

Alzheimer's disease is a progressive neurodegenerative disorder that affects large populations worldwide and produces measurable alterations in brain electrical activity. In this study, we propose an automated EEG based methodology for distinguishing Alzheimer's disease, mild cognitive impairment and healthy control subjects. The EEG signals are first decomposed using adaptive variational mode decomposition to obtain narrow band intrinsic modes that capture key non-linear and non-stationary properties of the recordings. A comprehensive set of time domain, spectral, Hilbert envelope and permutation entropy features is then extracted from each mode, resulting in an initial feature pool of fifty six descriptors. These features are reduced using the rich and poor optimization algorithm, and their discriminative ability is assessed using the Kruskal-Wallis statistical test with Bonferroni correction applied to control the family-wise error rate across all fifty six features (corrected threshold α = 0.05/56 = 0.000893). Forty-five statistically significant features are retained after Bonferroni correction and used as the final input feature set for classification. Three classical classifiers, namely random forest (RF), K-nearest neighbor (KNN) and support vector machine (SVM), are trained using tenfold cross validation with synthetic minority oversampling (SMOTE) applied strictly within each training fold to prevent data leakage in order to evaluate the proposed framework. The RF model achieves the highest performance with a classification accuracy of 82.10 percent, a balanced accuracy of 71.90 percent, a macro F1-score of 0.6534, and a macro area under the ROC curve (AUC) of 0.8933, supported by high sensitivity, specificity and predictive values. Cross-dataset validation on an independent publicly available EEG dataset (OpenNeuro ds004504, 36 AD and 29 healthy controls) yields an RF accuracy of 81.60 percent and AUC of 0.8922, confirming the generalizability of the proposed approach. The results demonstrate that the proposed combination of adaptive decomposition, multi-domain feature extraction and optimization driven feature selection provides an effective representation of EEG changes associated with cognitive decline. The methodology is computationally efficient and suitable for extension toward automated decision support tools for early detection of Alzheimer's disease.