Dynamically Enhanced LSTM Framework for Diabetes Prediction With SMOTE-Based Balancing and Grid-Optimized Hyperparameters

Diabetes mellitus is a chronic metabolic disorder that requires accurate and timely prediction for early diagnosis and intervention. Traditional machine learning methods often struggle to capture nonlinear interactions and temporal dependencies in clinical datasets. To address this, we propose a hybrid framework that integrates Dynamic Mode Decomposition (DMD) for feature augmentation with a bi-layer Long Short-Term Memory (LSTM) network for robust classification. DMD extracts latent dynamic patterns from static clinical attributes, transforming them into pseudo-sequential representations that the LSTM models use to capture long-term dependencies. The framework incorporates SMOTE-based class balancing and grid-optimized hyperparameters to improve robustness and generalization. Evaluation on the PIMA Indian Diabetes dataset shows that the proposed LSTM+DMD model outperforms standalone architectures, achieving 98.92% accuracy, 99.33% precision, 97.37% recall, and an F1-score of 98.34%. Ablation studies confirm the complementary roles of DMD and LSTM, while comparisons with state-of-the-art methods highlight the model's novelty and effectiveness. These results demonstrate that the proposed framework provides a highly accurate and interpretable solution for diabetes prediction, with promising applicability in clinical decision support systems.