<p>Diabetes remains a major global health issue and is rising rapidly. If left untreated, it can cause serious complications such as heart disease and stroke. However, inadequate patient data due to privacy and safety constraints often impair the model performance. Accurate blood glucose prediction is crucial for prompt action. To tackle this issue, a novel data augmentation method is proposed for tabular diabetic datasets called Feature Importance Guided MixUp (FIMU). FIMU uses feature importance scores, which were inspired by MixUp and k-MixUp, to produce a variety of synthetic samples. Five personalized diabetic patient datasets, along with the Pima Indian Diabetes Dataset and the Multiclass Diabetes Dataset, were used to assess the approach. Root Mean Square Error (RMSE), Mean Absolute Error (MAE), R² error and Clarke Error Grid (CEG) were used to evaluate the performance of four Machine Learning (SVR, RF, XGBoost, LightGBM) and two Deep Learning models (MLP, TabNet) trained on the augmented data. For Personalized datasets the experimental results show that FIMU significantly improves predictive accuracy, achieving RMSE values as low as 4.86, MAE of 2.58, and R² up to 0.97. For population-based datasets, RMSE values as low as 0.36, MAE of 0.19, and R² up to 0.98. In both the cases FIMU surpasses both standard MixUp and KNN-MixUp methods. The findings highlight FIMU’s potential for enhancing diagnostic reliability in data-constrained medical conditions by showing that it considerably improves glucose prediction accuracy.</p>