From EEG Classification to Personalized Neurofeedback: An Explainable AI Framework for Alzheimer's Disease

This study developed an explainable machine learning (ML) framework for Alzheimer's disease classification using resting-state EEG-BCI recordings from the OpenNeuro ds004504 cohort, which consists of 36 patients with AD and 29 Healthy Control (HC). Four EEG biomarkers and a normative-deviation score are integrated with feature weighting and logistic classification, while SHAP, gradient saliency, and layer-wise relevance propagation provide complementary explanations of model predictions. The performance of the framework was evaluated using stratified 5-fold cross-validation and a locked 15-subject holdout set (8 AD, 7 HC). The model achieved a mean cross-validated area under the receiver operating characteristic curve (ROC-AUC) of 0.823 ± 0.095 and a holdout ROC-AUC of 0.929. The Explainable AI (XAI) feature attribution identified posterior alpha power, alpha peak frequency, and frontal theta/beta ratio as significant discriminators between AD and HC (p < 0.05), with posterior alpha power showing the strongest association with classification (p < 0.001). Posterior alpha power alone achieved a ROC-AUC of 0.900 ± 0.074. These results demonstrate the feasibility of using explainable AI to extract important spectral features that can distinguish AD from healthy controls and may inform the future development of adaptive neurofeedback protocols tailored to individual brain patterns. However, a larger and more diverse data set is needed for future studies of real-time, personalized neurofeedback applications.