Gestational Diabetes Mellitus (GDM) must be anticipated early and precisely to prevent the complications to both the mother and the child. However, the prediction of GDM is most often done on an unbalanced, noisy, and heterogeneous clinical data, thus it is rather difficult to GDM predictions with a traditional machine learning model. In this study, a new preprocessing step is the suggested method which would enhance predictive capabilities of GDM since it would entail a multi-phase refinement of the data. Firstly, there are missing values that are addressed using Multiple Imputation by Chained Equations (MICE) and outliers using Interquartile Range (IQR) approach then, Z standardization is used to make data normal. The data is fixed by encoding categorical variables into one-hot so that they can be modeled. To address the issue of excessive class imbalance and noise In data A hybrid approach that uses SMOTE with Cluster Filtering (SMOTE-CF) and Weighted Edited Nearest Neighbor (ENN) is suggested. The SMOTE-CF is employed to generate high quality synthetic minority samples with pure clusters as opposed to the Weighted ENN, which removes mislabeled or borderline cases with a distance-weighted disagreement score. This is then cleaned up data that is being trained on a Gradient Boosting Classifier. The results of the experiments with two real datasets indicate that the recalls and F1-score improved significantly compared to the traditional preprocessing methods that indicates the clinical usefulness of the method in the timely and consistent diagnosis of GDM.