<p>The prediction of blood glucose (BG) plays a crucial role in diabetes self-management. Several machine learning (ML) methods have been proposed in the literature to create and evaluate well-performing models for the monitoring and prediction of blood glucose levels. However, single models are not always able to capture the inter- and intra-patients’ variations due to the complex nature of glucose dynamics. In contrast, ensemble learning, through the combination of multiple single learners, has demonstrated promising results in multiple medical fields including diabetology. The aim of the present paper is to build and assess the performance of deep homogeneous ensembles in predicting blood glucose levels. A novel approach to construct blood glucose prediction ensembles based on the Scott–Knott algorithm has been proposed. The study assesses how the performance of the constructed ensembles compares to that of their underlying single learners and investigates if a particular constructed ensemble outperforms the others. We conducted an empirical study with 100/60 variants of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), respectively, that were assessed using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) as statistical measures and Clarke Error Grid Analysis (CEGA) as clinical criterion. These variants were clustered using the Scott–Knott test and ranked using Borda count. The top ten ranked variants were used to create and evaluate the homogeneous ensembles. The results show that LSTM ensembles outperform their single variants. In contrast, CNN ensembles demonstrate similar performance to their individual variants. Furthermore, the constructed LSTM ensembles have been found to achieve better performance than CNN ensembles.</p>