Electroencephalography (EEG) is a non-invasive technique that records brain activity with high temporal resolution and is cost-effective compared to other neuroimaging methods. It can detect subtle neural activity changes before clinical symptoms appear, enabling early intervention and treatment. This study aims to develop deep learning models capable of classifying EEG recordings from healthy individuals, Alzheimer's disease (AD) patients, and frontotemporal dementia (FTD) patients. EEG data and participant metadata (age, gender, cognitive scores) are integrated to enhance model prediction accuracy. Three approaches are evaluated: a Multilayer Perceptron (MLP) trained on power spectral density (PSD) features, a ResNet1D (One-Dimensional Residual Network), and a custom convolutional neural network (CNN) with an attention mechanism. The models were evaluated using leave-one-participant out (LOPO) cross-validation. ResNet1D achieved the highest accuracy (84.41%), outperforming the MLP with feature extraction (76.14%) and CNN with attention (83.35%), with all models demonstrating balanced precision and recall.