This study introduces DECA-DiaXEL, a robust and interpretable machine learning framework designed for the early prediction of diabetes using high-dimensional clinical information and data from the WiDS 2021 dataset. At the core of this framework lies the novel Dimensional Expansion–Contraction Architecture (DECA), which first expands the feature space through domain-informed engineering and subsequently contracts it using a hybrid feature selection strategy. The expansion phase generates over 4,000 features capturing physiological variability, comorbidity indicators, temporal dynamics, and group-level statistics. The contraction phase mitigates redundancy and overfitting through a layered selection approach combining variance thresholding, progressive correlation elimination, cumulative importance filtering, and LightGBM-based ranking. DECA-DiaXEL also incorporates LightGBM-based iterative imputation to handle missing values and logical consistency enforcement during preprocessing. The final ensemble model, trained through extensive hyperparameter tuning and evaluated using ROC-AUC, accuracy, and weighted F1-score, achieved a ROC-AUC of 0.8781 and a weighted F1-score of 0.84, surpassing existing state-of-the-art benchmarks. Comprehensive ablation studies and SHAP-based interpretability analysis confirmed the robustness and clinical relevance of the proposed architecture.