Diabetes mellitus is a chronic metabolic disorder that demands early and accurate detection to prevent life-threatening complications. Traditional diagnostic procedures, such as blood glucose and oral glucose tolerance tests, are often invasive, time-consuming, and resource-intensive, making them less practical for widespread screening. This study explores the potential of artificial intelligence, specifically Feedforward Neural Networks (FNN), in predicting diabetes based on clinical data from the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). The main contribution of this research lies in the integration of the Adaptive Moment Estimation (Adam) optimization algorithm and the Synthetic Minority Oversampling Technique (SMOTE) to enhance the performance and generalization of FNNs on imbalanced medical datasets. The methodology involves preprocessing steps such as imputing zero values with feature means, normalizing input features using Min-Max scaling, and applying SMOTE to balance class distribution. Two model configurations were compared: a baseline FNN trained manually using full-batch gradient descent and an FNN optimized with Adam. Experimental results demonstrated that the baseline model achieved an accuracy of 70.13%, precision of 56.06%, recall of 68.52%, and an F1-score of 61.67%. In contrast, the Adam-optimized model achieved superior results with an average accuracy of 73.31%, precision of 60.97%, recall of 66.67%, and an F1-score of 63.64% across ten independent runs. These findings indicate that combining adaptive optimization with oversampling significantly enhances the robustness and reliability of neural networks in medical classification tasks. In conclusion, the proposed method provides a practical framework for AI-assisted early diabetes detection and highlights opportunities for future development using deeper network architectures and explainable AI models for clinical applications.