Diabetes mellitus is a widespread metabolic disease that requires continuous monitoring of blood glucose levels to prevent life-threatening complications such as hypoglycemia and hyperglycemia. Accurate prediction of blood glucose levels is crucial for improving diabetes management and patient outcomes. This research explores the application of deep learning models to forecast blood glucose levels using the OhioT1DM dataset. A key focus of this study is the optimization of data preprocessing, feature engineering, and hyperparameter tuning to enhance prediction accuracy. Various advanced time series models, including LightTS, Pyraformer, and Crossformer, are evaluated for their ability to predict glucose trends. Our results indicate that LightTS, when trained with a diverse set of additional features besides blood glucose, e.g. air temperature, heart rate, and physical activity, yields the lowest root mean square error (RMSE), surpassing previous state-of-the-art in the public Blood Glucose Level Prediction (BGLP) Challenge. By leveraging attention mechanisms and time-series forecasting techniques, this research demonstrates the potential of these models in medical applications beyond natural language processing. Our findings contribute to the ongoing effort to integrate artificial intelligence into healthcare, providing a foundation for real-time, personalized diabetes management.