The hallmark of Diabetes Mellitus (DM) is that the blood glucose levels are affected by insulin resistance, which characterizes this chronic metabolic disorder. Existing research often faces challenges of inadequate generalization, overfitting, and data imbalance. To address these challenges, a more accurate and reliable DM classification mechanism is proposed using LBSLLM. The proposed LBSLLM is an amalgamation of LSTM, SAM, and LEA. The alignment of missing values, scaling features for uniformity, and data cleaning to remove inconsistencies are incorporated into data preprocessing. The Gaussian Localized Weighted Synthetic Minority Oversampling with Distance Constraints (GLoW SMOTE-D) approach is proposed to address data imbalance, followed by more accurate DM classification using the proposed LBSLLM. The critical clinical features are emphasized, and temporal dependencies are captured using SAM with a lightweight Bi-LSTM. Robustness and classification accuracy, which are then enhanced using the lotus effect. Compared to the state-of-the-art, the proposed method achieves outstanding results with 99.65% accuracy, 99.49% precision, 99.07% recall, 99.52% specificity, and an F1- score of 99.47%, thereby improving predictive performance and providing a flexible, scalable solution for DM classification.
Eldho et al. (Fri,) studied this question.