EEGNetv4

As dementia prevalence rises with global aging, this study proposes a lightweight, privacy-preserving EEG-based diagnostic framework using deep learning and Federated Learning (FL) to address data scarcity, variability, and privacy concerns. Among five evaluated CNN models, EEGNetv4 achieved the best performance, reaching 97.1% accuracy with only 1,609 parameters and under 1 MB memory after hybrid fusion, while FL implementation maintained comparable performance at 96.9% accuracy. The results demonstrate a scalable, efficient, and privacy-compliant solution suitable for real-world clinical and edge-device deployment.