Abstract
Alzheimer's Disease (AD) is a very common neurodegenerative disorders and early detection using electroencephalography (EEG) can enable timely intervention, however, existing computational models often lack robustness, interpretability, and clinical scalability. This study proposes NeuroFusionNet, a hybrid deep learning framework for accurate, explainable, and efficient EEG-based classification of Alzheimer's Disease and related dementias. The model fuses handcrafted spectral, statistical, wavelet, and entropy features with latent temporal embeddings extracted from a customized one-dimensional convolutional neural network (1D-CNN). Feature selection is performed using Pearson Correlation Coefficient (PCC) and Particle Swarm Optimization (PSO), Principal Component Analysis (PCA)-based dimensionality reduction, and SMOTE-based class balancing has been performed to enhance discriminative learning. Comprehensive preprocessing including bandpass filtering, Artifact Subspace Reconstruction (ASR), and Independent Component Analysis (ICA) improves signal quality prior to classification through a five-layer deep neural network optimized via adaptive learning rate scheduling. Proposed method has been validated on three public EEG datasets including OpenNeuro ds004504 (eyes-closed), ds006036 (eyes-open), and the independent OSF dataset. Our method demonstrates state-of-the-art accuracy and macro F-1 score of 94.27% and 0.94 respectively. Cross-validation yielded minimal variance (SD <0.3%) that confirms the robustness and reproducibility. Model interpretability was ensured using Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), which revealed physiologically consistent biomarkers such as posterior alpha attenuation and frontal-theta enhancement patterns well aligned with established AD pathophysiology. Demographic fairness analysis showed negligible bias (<0.6% difference) across gender and age subgroups. Despite its high accuracy, NeuroFusionNet remains lightweight (0.94M parameters, 4.1 MB footprint) and computationally efficient (6.5 ms inference per sample), enabling real-time deployment on standard clinical CPUs without GPU support.