Gestational Diabetes Mellitus (GDM) is a relatively frequent complication of pregnancy with severe health effects of the mother and the fetus. Pre-diagnosis determination of GDM using the clinical data in the initial stages could be utilized to deliver effective early treatments, however, the misrepresentation of irrelevant and redundant features and noisy features complicate issues of machine learning models. To overcome these issues, this paper proposes very novel feature selection scheme named Hybrid Shapley-correlation filter with Adaptive weighting (HySCAF). The method proposed will combine the Shapley value-based feature importance with the model-aware evaluation, correlation-based filtering to get rid of redundancy, and an adaptive weighting mechanism which combines statistical variability and clinical relevance to select features according to their priorities. Applying a Kaggle publicly available GDM dataset, the success of HySCAF was confirmed by using it in conjunction with an XGBoost classifier. The model performance was assessed with the help of accuracy (87.4%), precision (85.3%), recall (89.1%) and F1-score (87.1%) and it was compared with other conventional methods like RFE, LASSO, Mutual Information and PCA feature selection methods. Findings indicate that HySCAF performed better than baseline methods in all metrics of predictive accuracy and interpretability. The framework presents a powerful and interpretable method of feature selection on medical datasets and has a high potential of being implemented in the real-life GDM screening systems. Future work involves testing with larger data sets, and incorporating domain knowledge in the clinical domain more systematically.